Skip to content

adobe/stardust

v0.39.4Apache-2.0

Redesign an existing website to make it better. Higher-level guided flow on top of impeccable.

Changelog

0.39.4 (2026-11-16)

Bug Fixes

  • stardust: a gate override waives only the criteria it names, and the unit cache knows its width (#704) (aac23ef)

0.39.3 (2026-11-15)

Bug Fixes

  • stardust: ledger.mjs names the readers for a guessed verb such as summary (#681) (facbc0b)
  • stardust: measure the page before a screenshot; a layout overflow is a WARN, not a crash (#687) (75e1d21)
  • stardust: mobile-first checks take a project's own breakpoint steps (#682) (b761eee)

0.39.2 (2026-11-15)

Bug Fixes

  • stardust: breakpoint-lint fails a missing --dirs file and negated desktop-first queries (#677) (8dda267)

0.14.0 to 0.39.1: entries written by hand

From here down, each entry was written with its pull request. The entries above are written by the release, from the titles of the pull requests; a pull request's description has what its change is about.

This part starts at 0.14.0. Prior versions (0.3.0 – 0.13.1) are documented in git history only (plus the branch-scoped notes in CHANGELOG-redesign-adobecom.md and CHANGELOG-delivery-media-fidelity.md).

0.39.1

  • Progress-reader guidance (#132). A progress reply suggested ledger.mjs summary, but the reader supports only tail and last. Skill instructions and the run-status reference now give existing commands: ledger transitions via tail / last [<skill>], page states via state.mjs summary. No CLI change; recorded status is not process liveness. Touches the ledger writer/start-guard guidance in 0.25.0 and 0.25.1.

0.39.0 — delivered CSS is mobile-first (#162)

The EDS code-review checklist flags any max-width query: base styles are the mobile layout, wider layouts sit in (width >= …) queries. Deploy lifted prototype CSS as written, and the prototype nav pattern, the deploy --nav-height example and the migrate / publish-sample audits all asked for max-width rules, so delivered sites failed that review. Replica already required mobile-first output (preserve-direction.md); nothing enforced it.

  • replica/scripts/breakpoint-lint.mjs: a direction check that runs with or without a target — any @media / @custom-media / @import upper bound (max-width, width </<=, the top of a range) fails with the mobile-first rewrite point (max-width: 599px → (width >= 600px)). media="…" attributes and JS thresholds are not direction-checked. No target set no longer skips the run.

  • Rule home: deploy Step 3 (#162) — convert desktop-first prototype CSS while lifting it; deploy wires lint:breakpoints into npm run lint on every project (replica C0 already did with a target). Pointers: deploy Lint bullet and checklist row; the --nav-height example (on the boilerplate's 600/900 steps) and the #109 note are now mobile-first.

  • Prototype: the Mobile-adapt audit refuses a desktop-first query (audit/responsive: desktop-first-query) instead of requiring a max-width rule ≤ 640px; the stock hamburger pattern and its fixture are mobile-first (@media (width >= 641px) restores the inline nav; same switch as before). Migrate and publish-sample point at the audit instead of restating it.

  • Folded: the #113 un-float trap (a mobile query un-floating columns does not occur in mobile-first CSS).

  • Tests: breakpoint-lint.test.mjs (direction on CSS only, rewrite point, no-target runs, prototype <style> vs <source media>). Nav fixture checked with mobile-nav-audit.mjs (clean at 360px) and at 641 / 1280px (inline nav, no burger).

  • deploy, prototype and migrate SKILL.md are outlines now. This change touches the three files, and each was over the repository's limits for a SKILL.md (500 lines, about 5,000 tokens; deploy had 1,189 lines and 203 KB). Every section keeps its heading there, so § pointers from other files still resolve, and names the file under reference/ that holds its text, moved as it was: 16 files for deploy (step-1-audit.md … step-10-reconcile.md, encode-contract.md, anti-patterns.md, checklist.md, …), 7 for prototype, 4 for migrate. No rule changed. A rule cited as #N is found by searching skills/deploy/ for it. player-before-paint-browser.test.mjs reads the #157 snippet from deploy/reference/step-3-foundation.md.

0.39.0 — one copy of the shared script rules, shipped with the master's scripts

Adding the EDS widget (0.34.0) touched 17 scripts, because every skill's scripts carried their own copy of the path, widget, chrome and kind rules. Those copies had drifted (nine path normalisers, five widget-stem helpers, two sets of class labels).

  • skills/stardust/scripts/lib/ (paths, names, kinds, taxonomy, resolve) behind contracts.mjs, next to browser.mjs. Setup step 7 already copies the master's scripts/ to stardust/scripts/stardust/, so the library ships with it, and a consumer finds contracts.mjs with the same two-layout lookup it uses for browser.mjs (0.38.0's mechanism; no new copy step). dynamics, qa, rollout and spec resolve it once in their lib.mjs.
  • Swapped, each script's own copy deleted: rollout (inventory, assemble, delivery-lint, lib, seed-redirects, da-library's page path), deploy (eds-path, deploy-batch, davids-model-lint, ai-readability), qa (lib, from-knowledge, checks), spec (lib, spec-knowledge, spec-map), migrate, dynamics, replica (gate-evidence, gate-all), diff (chrome-compare). Outputs are unchanged (golden snapshots byte-identical, field replays identical) except for these bugs, converged on purpose: a widget stem takes .htm and any case of .html, and, as deploy's block-roundtrip always did, a /widgets/… href with another extension or none (so delivery-lint and qa now read /widgets/x.js or /widgets/x as a widget link, where they required .html); only a /widgets/… href is a widget (rollout pageWidgets, qa widgetLinks and migrate widgetsIn took /a/b.html or /nav for one); spec-knowledge's class labels for A and D now read as dynamics' do.
  • Project copies: a stardust/scripts/stardust/ copied before this release has no contracts.mjs, and the scripts that need it stop with the copy instruction: re-copy the master's scripts. deploy/scripts/eds-path.mjs and spec lib.mjs loadSibling stay as re-exports for one release.
  • Tests: lib.test.mjs, layout.test.mjs (every skill copied the way the copy steps do, every script's --help run from the copy, a copy without the master's scripts stops with the instruction); the characterization test pins 106 divergences (103 − 10 converged + 13 for code added in 0.35–0.38).

0.38.6 — the scripts index lists what an agent runs, not the libraries it imports

The scripts index has a 12 KB cap so an agent can read it at session start, and it had sat within 10 bytes of that cap for eight releases: every new script meant shortening someone else's row. 17 of its rows (about 1 KB) were pure libraries, marked "no CLI", which an agent never runs.

  • evals/lint/libraries.mjs is the one list of pure libraries (exported functions, no argv), moved from script-help.mjs's exemption list. script-help.mjs still skips them for --help; scripts-index.mjs now requires no row for them and fails on a row that names one.
  • scripts-index.md drops the 17 library rows (12288 → 11282 bytes). The libraries an agent imports (browser.mjs, live-session.mjs, eds-path.mjs, measure-live.mjs, martech.mjs) are documented in the skills that use them, unchanged.
  • AGENTS.md: a new library goes in libraries.mjs, not the index.

0.38.5 — martech tags load from the boilerplate's consent phase

aem-boilerplate replaced delayed.js with a consent phase: loadDelayed() imports consent-check.js, whose placeholder hasConsent() declines by default and loads consented.js only on consent. The EDS building-blocks and code-review skills now state the rule (martech in consented.js). On such a project the scaffold hooked loadTags() straight into loadDelayed(), outside that phase.

  • deploy/scripts/martech-scaffold.mjs: when scripts/consented.js and scripts/consent-check.js exist, loadTags() hooks into consented.js, and the boilerplate's placeholder hasConsent() (with its JSDoc) is replaced by one that returns martech.js hasConsent(): true on the production hosts; off them ?consent=accept|decline decides as in the boilerplate, else ?martech=on. Each route keeps its own CMP category gate, so the source's consent model (per-tag, google-consent-mode) is unchanged. Only the placeholder, matched token for token, is replaced. Any other hasConsent() (an owner's CMP wiring, an async or arrow one, a placeholder reworded upstream) is a conflict that --force does not override, until it calls martechConsent(); then it is kept as is. A loadTags() hook an older run left in loadDelayed() or delayed.js is removed, so an upgraded site loads its tags once. Older projects keep the delayed.js / loadDelayed() hooks.
  • Docs: dynamics/reference/martech.md (runtime step, ?consent= switches).
  • Tests: martech-scaffold.test.mjs 24/24 — the consented.js hooks and re-run, the upgrade from an older run's hook, every non-placeholder hasConsent() as a conflict, an owner-wired one kept, the JSDoc, the hasConsent() matrix, and consent-check.js → consented.js → tags end to end.

0.38.3 — spec: classic AEM components inside responsive grids

Field learning from a medical-device manufacturer's run: the site mixes classic c-* components with responsive-grid members inside experience fragments. The aem-classic profile looked through the grids, so side-by-side members never became rows (the columns and cards a row maps to were missed) and components without the c- prefix went unnamed; the run had to extend the profile in its project copy.

  • aem-classic reads responsive grids: a grid member (aem-GridColumn) is the root, named by the classic component it wraps, so one component keeps one name inside and outside grids. Member widths group members into rows, as in aem-core, and the wrapped component is not counted as a second root.
  • parser.extraRoots (aem-classic) lists component classes without the c- prefix. The run needed four.
  • Pages without grids parse as before. On pages with grids, path ids change and grid containers appear as container: a run already past S7 re-runs parse and capture, and adds container to the mapping's layout list.
  • Re-parsing that run's 2,883 cached pages with this code and its four extraRoots gives output identical to the run's own parse. spec-parse.test.mjs covers grid rows, member naming and extraRoots.

0.38.2 — an mp4 goes to DA and is delivered on its own, authored root-relative (#161, amends #103)

Field learning (a migrated site whose hero carousel played three videos): the videos were committed to the code repo per #103. DA's Layout view (new authoring) renders the page from main--<site>--<org>.preview.da.live, which serves block code but 404s a code-bus mp4, so the hero showed blank there; authors also could not swap a video without a code change. A scratch-site test settled the DA route: a content.da.live/....mp4 link still reaches the page verbatim and answers 401 to visitors (#103 holds for that form), but previewing and publishing the uploaded mp4 itself makes its root-relative content path redirect to the media bus on both aem.page and aem.live (200 video/mp4, byte ranges served).

  • Rule (deploy/SKILL.md #161, replaces the #103 paragraph): an mp4 up to 36 MB (the content-source limit) goes through da-media-upload.mjs and is authored as /media/<scope>/<file>.mp4, never its content.da.live URL, and never also committed at the same path. Bigger mp4s, other formats, audio and PDF stay on the code origin or an external host. Pointer from replica's handoff contract.
  • deploy/scripts/da-media-upload.mjs: accepts .mp4 (video/mp4); after the PUT it POSTs preview and, unless --no-publish, live for the file on --branch (default main; --hlx-url for a test double), prints the root-relative path to author, and records authorUrl + delivered in the ledger. A failed preview/publish is FAIL (uploaded, not delivered), and a re-run redelivers it; an mp4 over 36 MB is SKIP video-too-large with no request. Manifest source fetches accept video/* for an mp4 entry. A video is stored under its web path (webPath: lowercase, runs of other characters → -), the only name its preview resolves (recorded: …_video01.mp4 uploaded fine and 404'd on preview); the OK line names the rename. Verified on a copy of the field site: after the move the hero plays in the Layout view and on aem.page.
  • Tests: da-media-upload.test.mjs (classify limits; PUT + preview + live order on a branch; failed preview; ledger skip of a delivered video; --no-publish; over-limit skip).
  • Review follow-ups: one skip path after the bytes are known, so an over-limit mp4 fetched from a manifest source is SKIP video-too-large too (a declared content-length skips it before the body is buffered); a 401 that persists on the video's preview/live POST halts the batch like the PUT 401 (exit 3, never failed); folder segments are kept verbatim (checked on a scratch site: only the file name takes its web form). Tests cover the three; the summary counts over-limit video apart from non-images.
  • Second review follow-up: two files that take one web name (Clip_A.mp4, clip-a.mp4) no longer overwrite each other on DA. The first in the batch keeps the path; a later one is FAIL name-collision before any request, with no ledger write (the row is the first file's), and a dry run shows it. Test included.

0.38.1 — rollout publishes the site's own DA block library (#160)

Field learning (a migrated site handed to its author): the pages and block code shipped, but the DA library showed an unrelated block set inherited from elsewhere, and none of the site's blocks or templates. Nothing in deploy or rollout wrote a library, so every migrated site arrived with an inherited or empty one.

  • rollout/scripts/da-library.mjs: from the delivered DA bodies (rollout coverage when present, else the content tree) writes library/blocks/<block> (the first authored instance of every class variant, each followed by a library-metadata description), library/templates/<template> (the template representative, its metadata as template-metadata with page-specific values emptied), the blocks / templates sheets, and merges the library tab of the site config: only the Blocks / Templates rows are replaced, other rows and tabs are kept. Previews the documents, never publishes them. --dry-run writes the files only.
  • Rule home: rollout/SKILL.md § Phase D (DA library + D-site checklist row); pointer from deploy § When you finish.
  • Tests: da-library.test.mjs (block and variant lifting, the documents in the shape the DA library parses, the config merge, the CLI dry run from coverage).
  • Review follow-up: a kept template metadata value with a $ pattern is copied verbatim (function replacement); the CLI runs through a symlinked path (realpath check, as in da-media-upload.mjs); a run without templates leaves the Templates row and templates.json alone; zero authored blocks exits 1 with no DA write; a 401 is retried before exit 3 and a network error is a FAIL (exit 1); a template whose representative is not delivered and a class with no blocks/<name>/ folder are named; content-tree chrome is the root nav / footer plus the documents pages name in their nav / footer metadata; template names use the label or unique id, never the shared type; author, dates and image-alt are emptied in templates; only template documents are previewed.

0.38.0 — every script takes its browser from one place (#159)

Scripts that imported Playwright, launched a browser, opened contexts and navigated on their own each picked their own launch mode, wait strategy and failure message, and a fix to one never reached the others.

  • skills/stardust/scripts/browser.mjs is the one entry to the browser: loadPlaywright, openBrowser (chrome, headed or plain), newContext, newPage and goto, plus everything live-session.mjs exports. live-session.mjs moved from skills/diff/scripts/ to sit beside it. Every script that touches a browser calls these by bare name. One error names the install line when Playwright is not importable, and a missing browser.mjs stops a script with the copy instruction. Where exit 1 means a failed check (render-harness, ai-readability, qa-gate, dynamics-check), both are exit 2.
  • A project copy needs the master scripts. Setup step 7 copies skills/stardust/scripts/ to stardust/scripts/stardust/ before any other skill's script runs. The skills that listed the diff scripts only for live-session.mjs (reskin, deploy) now name the master copy instead; replica still copies the diff scripts for the gates that import them.
  • A script the agent writes uses it too. Field check (a recorded hands-off replica run on this version): four scripts the agent wrote opened the source with chromium.launch(). The rule, with a recipe that works for a live page and a local one, moved to Setup step 7, which every run reads; replica and dynamics point to it. It names newLiveContext and gotoLive: in a second run two scripts took the browser from browser.mjs and then called page.goto themselves, which sends HeadlessChrome in the user agent. ledger.mjs … end names, in one warning, the scripts written in that phase that import Playwright themselves.
  • No longer works: a script copied without the master scripts. A script copied flat into stardust/scripts/ (the crawler, the style census, the dynamics helpers) still runs when they are in stardust/scripts/stardust/.
  • A script run from the plugin tree finds Playwright in the project at the working directory first, then by import('playwright'). The first copy that resolves is the one that loads. A copy that fails to load stops the script with its cause (a missing dependency included); it is never replaced by the other copy.
  • No script changed its browser. Each keeps its launch mode, launch options, context options and navigation options.
  • evals/lint/browser-front-door.mjs fails on a script under skills/ that imports Playwright (from, import, import(), require(), resolve(), playwright or playwright-core), launches or connects a browser, opens a context or browser.newPage, or calls .goto( itself. It runs in npm run lint:stardust and in CI. A skill argument that is a file, not a skill folder, exits 2 like an unknown name.
  • playwright-cli is no longer listed as a requirement (compatibility: lines, README Prerequisites): no script uses it.

0.37.2 — a named page count starts from the entered page and its links (#158)

User feedback: asked for N pages of a site, the crawl filled the slots in sitemap order, so the pages the entered URL links to were often cut.

  • extract/scripts/crawl.mjs --link-first: orders the pages entry first, then the entry page's same-origin links in document order, then the other discovered pages, up to --max. Only a user-named N sets it; --max alone (default or roster caps), --max 0 and --pages are unchanged. In link-first mode both lists keep same-origin non-asset pages only (ASSET_EXT). An unparsable href is skipped with a WARN instead of aborting discovery, in both modes. Rule home: ia-extraction.md § Priority for the cap; pointers from extract Phase 1, --prep (a user-named N keeps --cap N instead of --all) and the hands-off volume caps (a user-named N replaces 100/20).
  • Tests: crawl.test.mjs parseArgs (--max alone never turns link-first on), selectPages (order, BFS re-entry, unparsable hrefs), discover on a stub page (sitemap + malformed anchor, BFS fallback).

0.37.1 — a player iframe mounts after first paint, and qa names the request that came before it (#157)

Field learning (recorded replica runs of one source page): a block in the first section created the source's YouTube iframe in decorate(), so the player's JS started while the section was still loading. Lab tests count that toward FCP/LCP: a page painting in about 1 s measured lab LCP up to 8.5 s, scores 52–100 run to run. lcp-poor saw the symptom, nothing named the cause.

  • Rule (deploy/SKILL.md #157 + checklist row): same src, later mount — an aspect-ratio box at decorate time that keeps the authored link (a player-only section still paints), replaced by the iframe after the first contentful paint (without paint timing: at load, or at once when already loaded), from an allow-listed https: URL with createElement; the iframe is display: block with no border, so the box keeps its height when it mounts. Pointers from replica's live-data embeds bullet (deferral noted in the progress ledger) and dynamics' media-as-url.
  • qa/scripts/checks/perf.mjs: player-before-paint (warn, never downgraded) lists player document, script and API requests from any frame that started before first contentful paint, dated by the browser: timing().startTime, or the CDP wallTime while a response is still pending. Without either, two bounds decide (heard before the paint; frame created after it); otherwise player-timing-unknown (warn), never a silent pass. Posters on player hosts are not flagged. Tests: qa/player-before-paint (pure judgePlayerRequests), qa/player-before-paint-browser (collector under a stalled event loop, a held-open player request, snippet lifecycle, untrusted links; Playwright from the project, routed, no network).
  • dynamics modal-loader example: the player iframe is built with createElement from a parsed https: URL on a dot-bounded host match, no longer interpolated into innerHTML.

0.37.0 — a resumed run keeps its gate verdicts, not its pictures (#156)

Field learning (a recorded 550-page site-scope replica run in a 20 GB container): the disk filled after every page was published. The tracked tree was 5.3 GiB, 82 % PNG: gate captures and diffs of every page (3.3 GiB, a fifth of them extra captures left beside a page's own) and extract screenshots (0.76 GiB). A harness that commits all of stardust/ so a restart finds the same files pays each byte twice (worktree and .git); it could not follow the ignore file instead, because that dropped whole folders a resumed run reads. Design note: notes/disk-lifecycle/README.md.

  • stardust/reference/stardust.gitignore: ignores gate images, not gate folders — **/gates/**/*.png (a frozen baseline.png stays; first-page/gates/ is covered too) — and tracks current/assets/ except its screenshots. On that run: 0.5 GiB tracked instead of 5.3 GiB. Existing projects keep their copy (Setup step 6 never edits it): replace its current/assets/ and replica/gates/ lines with the two gate lines to adopt it.
  • replica/scripts/gate-all.mjs: a page that passes keeps its records, not its PNGs. Every PNG in its dir is deleted (an extra capture too) and pixel.json records the drop with the masks and eds URL it was measured with; --skip-existing and --compare-only reuse that pass without a capture while both hold. A failed or override-delivered page keeps its captures for the fix round. Pages go through in batches of 25, so a passed batch's PNGs are gone before the next batch is captured (peak disk). A page captured again loses its recorded pass first, and a missing capture is written to pixel.json: a failed or interrupted re-capture never leaves the old pass to be reused.
  • replica/scripts/gate.sh: a round deletes earlier rounds' diff-*.png and vdiff-*/ before comparing; their text and JSON evidence stay.
  • Docs: artifact-map.md § Versioning rows, stardust/SKILL.md § Versioning, the state report consequences in state-machine.md, source-fidelity-gate.md § all-pages gate, the README; replica/SKILL.md's copy of the delivery-gate file list folds to a pointer at gate-all.mjs --help.
  • Tests: gate-all.test.mjs — dropPassedCaptures, droppedPass, and runs with stub instruments (a passed page loses its PNGs, a failed one keeps them; reuse without a capture; a changed mask re-captures; a failed or interrupted re-capture captures again on resume; at most one batch of captures on disk at a time); gate.test.mjs — a round keeps only its own diff image and vdiff dir.

0.36.7 — the site secret: exact-origin match, never context-wide

The site secret (--auth-header / --token-env, default SITE_TOKEN) is matched against the exact site origin (parsed origin, not a string prefix) and never set context-wide: the AI-readability gate (deploy/scripts/ai-readability.mjs and the qa check) attaches it through the origin-scoped route filter instead of extraHTTPHeaders, and its raw page and fragment requests drop it once a redirect leaves the origin. The qa AI-readability check and scorePage skip paths that resolve off the origin, and dynamics-plan.mjs --target-origin probes only API paths on the target origin.

Anyone who ran qa or ai-readability.mjs with a site secret on 0.20.0–0.36.6, or an earlier source snapshot containing site auth, should rotate it. Tests: deploy/scripts/test/ai-readability.test.mjs, ai-readability-redirect.test.mjs, qa/scripts/test/origin-auth.test.mjs, ai-readability-auth.test.mjs, dynamics/scripts/test/dynamics-plan.test.mjs.

0.36.6 — one patient visitor: shared live pacing adapts and the browser keeps its cookies between instruments (#155)

Field learning (recorded hands-off replica runs): strict origins rejected clustered live navigations while friendly origins did not; every instrument opened a fresh browser context, so each capture looked like a new visitor that had never cleared a challenge or accepted consent; and parallel workers each ran their own instruments against the same host. Fixed pacing was either too slow or not cautious enough.

  • Adaptive shared throttle (diff/scripts/live-session.mjs): fresh hosts allow STARDUST_LIVE_MAX_PARALLEL (default 2) in flight and start at STARDUST_LIVE_PACE_MS (default 1000) between starts. A shared bot signal reduces the host to one navigation at a time and doubles the pace through 2000, 4000, 8000 and 15000 ms. Signals are an edge challenge, a denial page, any 429 (Retry-After honoured up to 120 s), a 503 with Retry-After, or a timeout or connection error after that host has answered. Every 50 clean navigations halves the pace down to the configured floor, then one more such window restores the configured parallelism, and a host's backoff is forgotten STARDUST_LIVE_STATE_TTL_MS after its last signal. Repository-scoped OS-temp state, fair queueing and stale-lock recovery coordinate every process. Lost waiters re-queue and stale waiters/leases expire; STARDUST_LIVE_MAX_WAIT_MS (default 300000; 0 disables) caps slot waits, even during a hold, without clearing it. 0 disables either throttle limit, not holds. Bookkeeping lock waits cap at 10 s; failures warn once per host, then navigation proceeds unthrottled; transport response marks prevent duplicate challenge accounting.
  • Per-host browser state: after a successful gotoLive the context's storage state is saved (atomic write, at most every few seconds per process) and newLiveContext loads it when it is younger than STARDUST_LIVE_STATE_TTL_MS (default 12 h) and the caller passed none, so a cleared challenge or an accepted consent carries over to the next instrument; save failures only warn, and STARDUST_LIVE_STATE=0 disables it. The state never lives inside the repository.
  • Every live instrument takes part: the crawler (its entry probe and a challenged page report the signal once each — the reloads that stay challenged are the same wall; a captured page counts as a clean navigation), census, gate probes, measure, parity, motion, anchor, sibling-variance and the reskin captures create their contexts through the shared helper (one-line changes), with inline no-op fallbacks where a project copy lacks the module. Tests: diff/live-session (throttle decisions with a fake clock and a temp dir, stale-lock reclaim, fairness, storage-state round trip), extract/crawl (fallbacks).

0.36.5 — the style census runs one live page at a time unless an operator allows more (#154)

Field learning (a recorded hands-off replica run on a bot-managed site): right after the crawl, the style census — a second full live pass at two widths — was started with --concurrency 2, although the option's default is 1 and its help text warned to raise it only on an origin known to have no bot manager. The origin stopped accepting connections and 17 of 20 census page loads timed out, which ended the run. The warning was advice; it is now a guard.

  • extract/scripts/style-census.mjs: --concurrency <n> above 1 is honoured only with the new explicit --parallel-live-ok, or when every input is local (file:, localhost and *.localhost, 127.0.0.1, [::1], 0.0.0.0); otherwise one stderr line (--concurrency 2 ignored — one live page at a time unless --parallel-live-ok …) and the census runs at 1.
  • Browser tier: when the shared live-session module beside the script exports launchLiveBrowser, the census launches through it (the identity-preserving default tier, paced per host); --headed keeps the shared escalation tier; a copy without the module keeps the plain headless launch. Tests: extract/style-census (+ guard, local exemption, help text, launcher fallback).

0.36.4 — chrome-explore paces itself and keeps a per-width request budget (#153)

Field learning (a hands-off replica run on a site behind an edge rate rule): the crawl, the style census and the cap probe went through; the header exploration then reloaded the page after every control (routing disables the browser cache, so each reload refetched the whole page) and typed the probe term keystroke by keystroke into the site's typeahead — the edge blocked the address mid-run.

  • diff/scripts/chrome-explore.mjs: --pace-ms 400 after each control open / close and --reload-pause-ms 1500 before each reload (0 disables either). --max-requests 600 per width counts every request leaving for the site's host — documents and subresources alike, the first load included; 204-answered navigations not — and ends the width with truncated: 'request-budget' and n/max requests on the summary line (a lower budget was spent by an image-heavy page's first load alone while recording the same contract). --max-typeahead 2 inputs per width; the term enters as one fill plus one typed letter and a single settle, inputs past the cap are listed with search: null.
  • Closing in place: between controls the header is closed the way it was opened — Escape and the parked pointer, each opened trigger re-acted innermost first, a click outside — and the page reloads only when atoms are not back as loaded (visibility and aria state, so a left-behind selected tab still reloads). A GET search submit is answered 204 like every navigation and its URL is the record; a POST is followed with the results page's assets aborted, then the page restored. The probe word is read from the explored page on first use (no extra load). Fixture headers, same recorded states: rich 26 → 13 loads at 1440, 24 → 1 at 390; stock 6 → 1 at 390.
  • Exports and docs: parseArgs, requestBudget, sameSite, siteHost are exported; --help documents the four flags; scripts-index.md names --max-requests.
  • Build exploration: diff and qa pass --max-requests 0 --pace-ms 0 --reload-pause-ms 0; source defaults stay protective.
  • Parity: request-budget gaps warn with the unbudgeted-build remedy; observed failures remain errors.
  • Tests: chrome-explore.test.mjs — the summary line, parseArgs defaults and fallbacks, sameSite / requestBudget; with a browser: closing in place, the GET and POST submits, a stubborn header (new fixture in chrome-fixtures.mjs), the budget stop and the typeahead cap.

0.36.3 — the crawler names a bot wall in seconds instead of waiting out its cap (#152)

Field learning (a hands-off replica run on a bot-managed site): the entry page answered 200 with a verification interstitial that reloaded itself every few seconds; the crawler captured it as content and waited out its whole cap on it. A CloudFront block on inner pages read as a plain HTTP error, and the headed escalation opened a window where no display exists.

  • extract/scripts/crawl.mjs: launchWithFallback launches real Chrome headless first — its own identity, no stealth flags; bundled Chromium when Chrome is missing (technique chrome | chromium-fallback) — through live-session's launchLiveBrowser when the module beside the script has it, inline otherwise. probeEntry classifies the entry page in seconds: a network reject (fingerprint), an edge challenge (challengeMarker: Cloudflare / Akamai / F5 / Imperva / CloudFront–AWS WAF, Akamai's 400 escalation) or a 2xx interstitial read from the page — the crawler's own title rule first (wallProbeInPage / inlineWallMarker: a wall title or h1 on a page under 2 KB of text or one reloading itself), then live-session's pageDenial vendor list when present. Any of them → one window-free real-Chrome relaunch (technique chrome-escalated; the existing launch flags and navigator.webdriver spoof stay on, STARDUST_STEALTH=0 turns them off; a relaunch failure is one stderr line) plus the challenge-solve window, then a fatal bot-management <kind> not cleared: <marker>. On inner pages an uncleared challenge or a 2xx interstitial is recorded as errorClass BotChallenge with the marker — never as content, never as a bare HTTPError; _crawl-log.json#discovery carries botBlock + botMarker.
  • The wall title names verification interstitials only — "verify you are human", "verifying your request / browser", "human verification", "checking your browser" and the vendor interstitial titles, never a bare "verify" or "verification": an age or e-mail verification page is a page, and a small single-page-app entry page under such a title would have aborted the crawl. Readers of fetchTechnique (deploy/SKILL.md, deploy/da-deploy-protocol.md, replica/reference/handoff-contract.md) know the tier names (chrome, chrome-escalated, chromium-fallback; older logs headed-chrome*) and verify source images with an in-page fetch from a browser of that tier, never a bare curl.
  • --page-timeout <s> (default 60; 0 = none) wraps each capture: a page that never settles is a TimeoutError and its tab is closed. Live navigations are paced per host when a live-session copy that paces is beside the script.
  • Docs: extract/SKILL.md § Bot-management block says what the crawler does itself; scripts-index.md lists the flag.
  • Tests: crawl.test.mjs — the classifiers, the inline wall rule, pageTimeoutMs, probeEntry and capturePage on stub pages (clean / interstitial / challenge / CloudFront / bare 403 / hung navigation / an unreadable wall probe).

0.36.2 — real Chrome identity, retained automation-signal handling and paced live navigation (#151)

Field learning (hands-off replica runs on bot-managed sites): the gate instruments presented a hard-coded desktop user agent on a Linux headless engine and were denied; the agents patched the identity by hand at run time, and spacing live navigations a few seconds apart turned the denials into 200s. The escalation to real Chrome opened a window and could not start where no display exists. The policy now has one home: every context presents the browser's user agent without the headless token and matching client hints; automation-signal handling stays on unless STARDUST_STEALTH=0; STARDUST_HEADED_WINDOW=1 opens a window.

  • diff/scripts/live-session.mjs: launchLiveBrowser is the default tier of every live instrument — the installed Chrome headless (channel: chrome), no stealth args, no ignoreDefaultArgs; bundled Chromium only when Chrome is missing (tier chromium-fallback, one stderr line). Every context presents the user agent read from the running browser with the headless token removed (HeadlessChrome/156.0.0.0 → Chrome/156.0.0.0), or the caller's --ua. Headless Chrome's header says HeadlessChrome while its client hints say Google Chrome, and managed challenges key on that token. The sec-ch-ua* client hints derive from browser.version() and the running OS on Chrome tiers, from the UA string under --ua, and stay untouched on the bundled tier. The existing navigator.webdriver spoof reaches every context. launchStealthHeaded (--headed) stays window-free real Chrome plus the challenge-solve window and the two existing launch options. STARDUST_STEALTH=0 disables the spoof and options with one stderr line; a window stays behind STARDUST_HEADED_WINDOW=1. gotoLive paces live navigations per host STARDUST_LIVE_PACE_MS apart (default 2000; 0 disables; slots reserved synchronously so concurrent callers queue; one stderr line per host on its first wait — an Akamai-fronted origin answered 403 to back-to-back navigations, 200 once spaced). Classifiers: a 403 with server: CloudFront or x-cache: Error from cloudfront, or any x-amzn-waf-action header, is the CloudFront / AWS WAF edge signature; F5's BigIP header matches; denial pages "The request could not be satisfied" + "Request blocked", "One moment, please…", "Please wait while your request is being verified".
  • Wired instruments: replica/scripts/stitch-shot.mjs, diff/scripts/content-diff.mjs, visual-diff.mjs, measure-live.mjs (openBrowser default tier chrome; stealth stays an alias of the --headed tier), replica/scripts/chrome-parity.mjs and motion-observe.mjs launch the default tier and default --ua to the browser's own; each --help says so. extract/scripts/style-census.mjs, replica/scripts/measure.mjs and cap-probe.mjs drop their hard-coded desktop user agent the same way — a container run showed them still presenting it beside real Chrome's own client hints.
  • Docs: source-fidelity-gate.md § Hardening rule 1 owns identity, retained automation-signal handling and pacing; rule 12's ladder, diff/SKILL.md (--ua, --headed) and replica/SKILL.md point to the explicit opt-out.
  • Tests: live-session.test.mjs (50 checks — automation defaults and opt-out, pacing, user-agent and identity derivation, client hints and denial classification) and stitch-shot.test.mjs (--ua defaults to the browser's own).

0.36.1 — replica: a capture names its dead font faces; a font check that could not run is never a pass (#150)

Field learning (a recorded hands-off replica run): the foundation CSS declared its metric fallback faces with a local() source naming a platform font the gate's Linux browser does not have. The build rendered a different face than the live site, the first archetype gate failed on type metrics, and the agent spent the better part of an hour finding the cause — the capture's font-load assertion reported only a face whose file failed to fetch, so an unresolved local() looked like a clean capture, and a check that threw was recorded as "every face loaded". In a later run the capture named the failed families and the fix took seconds: locally available metric-compatible faces (Liberation Sans / Arimo, Liberation Serif / Tinos) as the local fallbacks.

  • stitch-shot.mjs: after document.fonts.ready every declared face with status error and no loaded face of the same family, weight and style is a dead face — a file that failed to fetch or a local() source this browser cannot resolve; a failed declaration beside a loaded twin is not, an unloaded face was never needed. The WARNING names the families and both causes; the capture record (<stem>.capture.json, 0.34.4) carries fontsFailed ([] = every face loaded; null + fontCheck when the check could not run, with one stderr line — never silently a pass). A --freeze-baseline capture whose check could not run is refused like one with dead faces.
  • gate.sh / gate-all.mjs: both read the capture records (<side>.capture.json; the live one lives and dies with live.png; under --gate-against-baseline the live side reads the frozen baseline's check). gate.sh prints one gate.sh: fonts on the <side> capture — … line per side after the overflow line — the dead faces, every face loaded, or NOT CHECKED — as information for the loop (exit codes unchanged). gate-all carries fontsFailed on the origin / eds meta (the prototype stage reads live.capture.json beside the reused archetype live.png, or the baseline's record after a flagged round; none is unchecked) and fails the row: font-not-loaded:<family>, font-check-skipped:<origin|build> (fontsNotLoaded, fontChecksSkipped; totals, the calibration line, FAIL (fonts)). A meta from before the check (no key, --skip-existing) is not flagged — recapture to check.
  • Fonts policy (recreation-procedure.md): self-host the source's own faces; brand family first, a metric-matched substitute only as the fallback entry after it; a local() source only for a face the gate's browser has, the Linux-available metric fallbacks named; proprietary faces raise deploy's licensing alert (#80). Gate rule 14 names fontsFailed and font-check-skipped. Tests: replica/stitch-shot (new), replica/gate, replica/gate-all.
  • Merged with 0.34.5. The branch's --json <file> record was folded into the capture record every stitch already writes since 0.34.4, so a capture has one record; one stitch-shot test file carries all four contracts.
  • Merged with 0.36.0. The capture record carries fontsFailed beside 0.34.12's fixedLayers; gate-all reads a reused capture's font check from the record it carries — the archetype round's live.capture.json, or the frozen baseline's after a flagged round — and a capture staleCapture retakes is checked afresh.

0.36.0 — first page first: one page is delivered while the site scope is still being defined (#149)

A recorded hands-off replica run ran every phase in sequence and published no content page in five hours, though almost none of the site-wide work changed the home page. Design note: notes/first-page-first/README.md.

  • replica/reference/first-page-first.md (new): a full-site run opens with two tracks, one writer per path. Track A takes the first page (replica Input --first-page <url|home>; hands-off: home) through the bounded entry to the published origin, in its own workspace stardust/first-page/, and owns code and DA; its Phase 2 is always the bounded branch (preserve-direction.md § 1). Track B runs Phase 1 --prep, the catalogs and the dynamics triage and writes knowledge only. At RECONCILE, Track B's design spec replaces the bounded one, then seven checks (page type, blocks, chrome, dynamics, listings, scope, canon) are recorded confirmed | corrected (<commit>) | deferred (<why>) in stardust/reconcile-first-page.md; a block or chrome correction re-gates the page. Hand-off to rollout: update-coverage.mjs --gate <table> --pin, then verify.mjs --slug.
  • Scripts: ledger.mjs adds replica phases first-page and reconcile-first-page (alias reconcile) and --track <name> (start/end pair per track; last <skill> --track); update-coverage.mjs --gate --pin matches rows by path, pins them and marks them deployed; inventory.mjs never flags a pinned row stale. Tests cover each.
  • replica/SKILL.md and stardust/SKILL.md § Hands-off defaults point to the new reference.

0.35.0 — the gate has two tiers: looks the same, nothing lost; owners accept changes in one file (#148)

Field learning (a recorded home-page replica gate): gate.sh --full failed on any structural 🔴, so intentional look-neutral changes failed forever, invisible losses were split between 🔴 and buried 🟡, and no instrument read the owner's accepted changes. A page without <main> was compared at <body> against the build's <main> (142 🔴 instead of 53), and gate records cited probe files by hand.

  • Two tiers (diff/scripts/content-diff.mjs): looks-same = every 🔴; nothing-lost = MISSING ARIA-LABEL (now 🟡 everywhere), MISSING SR-ONLY TEXT, ROLE DOWNGRADE; the rest is advisory. A Tiers: looks-same 🔴 N · nothing-lost 🟡 L · accepted 🟢 A (owner O, auto P) · advisory M line follows Findings: (unchanged); auto = ROLE UPGRADE / SAME IMAGE / SAME ICON; --json findings carry their tier.
  • Accepted list: stardust/replica/accepted.json (or --accepted <file>), one owner decision and reason per entry, read by content-diff (applyAccepted) and chrome-parity (markAccepted): a match prints 🟢 and never counts. text matches inside the finding's quoted subject (subjectOf), never its wording; an entry without text accepts every finding of its kind.
  • Aligned roots (alignPrimary): when only one side's main root fell back, the other is re-read where that fallback really lands (no <main> → both at <body>), said once in a ⚠ root "main": … line. inventory() no longer counts its own width probe as copy at a <body> root (diff/ and deploy/ content-inventory.mjs).
  • replica/scripts/gate.sh: prints the Tiers: line on its content-diff: verdict line, fails the round on a nothing-lost 🟡 as on a 🔴, and writes verdict-<label>.json per --full round (every probe's exit code and evidence file, both tiers). gate-evidence.mjs passes content-count only at 0 🔴 and 0 nothing-lost, both from the last round (parseFindings).
  • Docs: source-fidelity-gate.md § Pass bar item 1 is the one home of the tiers, accepted list and verdict record; replica/SKILL.md, diff/SKILL.md and handoff-contract.md row 4 point to it.
  • Tests: content-diff, chrome-parity, gate and gate-evidence .test.mjs cover each change, incl. a no-<main> page against a <main> build.

0.34.14 — diff-report names the block a pixel difference belongs to (#147)

Field learning (a recorded replica gate on a home page): the band table says where drift sits, never which block, and the diff image is 8–15k px tall; block attribution and crops were built ad hoc.

  • replica/scripts/diff-report.mjs (new): reads a gate dir's live.png, build.png and diff PNG (the newest diff-*.png, or --diff), loads the build at the round's width and reads its block boxes ([data-block-name] plus each .default-content-wrapper as "default content (section N)", else main > .section, else main's children, or --blocks <css>; nested blocks fold into the outermost). Prints the differing share per block and outside every block. When live and build heights differ it warns; a block below the compared height reads "not compared" (never a false 0 %), one straddling it "partial". Writes report.html, a live | build | diff overview.png and crops of the --top hottest blocks.
  • Docs: source-fidelity-gate.md § Reading the band breakdown points to it; scripts-index.md gains its row.
  • Tests (new): replica/scripts/test/diff-report.test.mjs — the pure parts, --help, and end-to-end gate dirs (a two-block page; an EDS-shaped page with default content, a nested block and a build taller than live).

0.34.13 — chrome-parity prints real deltas only: an auto-detected header, painted backgrounds, icons by pixels, one line for a shifted region (#146)

Field learnings (two recorded replica gates): on a home page without <header>, chrome-parity skipped the header and printed 24 invisible footer deltas (display: contents titles, re-hosted thumbnails, an <img> logo, a background painted on another element). On an article page, 67 deltas were one fact: the footer sat 3,526 px lower as a block.

  • replica/scripts/chrome-parity.mjs:
    • the default header region on a page without <header> is auto-detected (the outermost full-width, top-anchored bar with a link, 30–400 px) and the region line names it; an explicit --region is never second-guessed;
    • text in a display: contents element is measured by its text box;
    • the region background is the painted one (the region's own, else a child covering ≥ 90 % of it, else the nearest ancestor's);
    • icons also pair by pixels (diff/scripts/image-signature.mjs): a re-hosted file or an svg drawn as <img> is no delta;
    • a tag change alone (same style, same box) is context on a real delta, never a delta;
    • a region whose box and every paired text (≥ 2) moved by one Δy (±2 px) is one advisory OFFSET line, outside the exit code; texts that moved inside a box that did not, or only some of them, stay deltas;
    • a --live-cache written by an older probe (PROBE_VERSION) is re-probed.
  • Docs: source-fidelity-gate.md § chrome-parity — the new behaviour in one sentence.
  • Tests (new): replica/scripts/test/chrome-parity.test.mjs — collapseOffset, compareRegion icons by pixels, --help, and an end-to-end page pair with every case above: only the advisory OFFSET remains, exit 0.

0.34.12 — stitch-shot captures fixed and sticky layers once: no seam repeats, no content hidden under them (#145)

Field learning (a recorded replica gate on a home page): both captures repeated the sticky header in every viewport chunk, covering whatever sat under each seam on both sides — an oversized partner logo under it scored 0.0 %, so a defect under a seam was invisible to the pixel gate. The same run stitched an undismissed consent bar into every live chunk unwarned.

  • replica/scripts/stitch-shot.mjs: fixed and sticky layers are captured once. Chunk 1 shows them; from chunk 2 a fixed layer, and a sticky one while stuck (displaced from its own in-flow place, top or bottom), is set to opacity: 0 — layout never moves. A sticky element still in its place mid-page stays. Layers are re-scanned after every scroll, so a header a scroll class makes fixed, or a widget injected later, is hidden from where it appears, with a stderr WARNING. A backdrop (negative z-index, or ≥ 90 % of the viewport) is never hidden. A fixed layer still visible after the dismissal sweeps, ≥ 5 % of the viewport, below the top bar and reading like consent / sign-up copy, is a stderr WARNING naming it and --consent / --dismiss. capture.json records fixedLayers: { capture: 'once', count, late }. Shadow-root, iframe and pseudo-element layers are not seen.
  • replica/scripts/gate.sh, replica/scripts/gate-all.mjs: a cached live.png (or a gate-all --skip-existing origin / eds capture) whose capture record has no fixedLayers is re-taken with one line, never compared with a build captured the new way (staleCapture). --freeze-baseline (0.34.5) records fixedLayers in baseline.json; a baseline frozen before this change reads as stale (re-freeze with --force), so --gate-against-baseline never compares the two capture methods.
  • Docs: recreation-procedure.md § Fixed and sticky chrome describes what the instrument does and does not cover; gate doc rule 11 and the replica SKILL pointer shortened.
  • Tests: stitch-shot.test.mjs — layerAction, isBackdrop, overlaySuspects, layerPassScript on a stub DOM, and end-to-end fixtures (fixed header, in-place sticky label, undismissable cookie bar, scroll-triggered header, backdrop); gate.test.mjs — a stale live capture is re-taken; gate-all.test.mjs — staleCapture.

0.34.11 — content-diff pairs icons by their pixels: a re-hosted logo or an svg drawn as an image is the same icon (#144)

Field learning (a recorded replica gate on a home page): 6 🔴 ICON DIFF were logos the build re-hosted under media_<hash> with ≤ 1.2 / 255 mean pixel difference, because identity was the file name; 5 🔴 MISSING ICON were 0×0 inline svgs or inline-svg logos the build draws as <img src="*.svg">.

  • diff/scripts/image-signature.mjs (new library): an icon's signature is its element screenshot averaged to a 12×12 RGB grid, so a recoloured shape is another image. sameImage = mean |Δ| ≤ 12 / 255 over the drawn cells with boxes within 2 px; a near-uniform grid (an empty box, an unrendered glyph) is never an identity.
  • content-diff.mjs: visible icons are signed; laid-out 0×0 icons are skipped (display:none ones still compare by name). Same pixels are the same icon when pairing anchors and ICON MOVED (two blank boxes never pair). A pair differing only by file or technique is accepted — SAME IMAGE / SAME ICON on the root's accepted line — never a finding; a different logo stays 🔴 ICON DIFF.
  • Docs: diff/SKILL.md § Reading content-diff names pixel identity; the deploy/SKILL.md Step-10 copy set names every local import.
  • Tests: new test/image-signature.test.mjs (blank, a recoloured shape, thin-line icons); content-diff.test.mjs — an empty box missing on the build stays MISSING ICON; end to end, a re-hosted file is SAME IMAGE, svg vs <img> pairs, a different logo stays 🔴, a 0×0 svg is skipped.

0.34.10 — content-diff severity follows visibility: a same-look heading is an improvement, not a swap (#143)

Field learning (a recorded replica gate on a home page): 23 structural 🔴 were ROLE SWAP for footer and mega-menu group titles — source <div>s, same-look build headings — while the one real loss, a chart's visually hidden summary, was a buried 🟡 MISSING BODY.

  • content-inventory.mjs (diff + deploy copies): each item carries its computed look (size, weight, colour, text-transform) and srOnly (sr-only clip, ≤1 px clipped box or -9999px offset on the element or its 3 nearest ancestors; display:none and off-screen slides excluded). With the same look and face, body/eyebrow → heading is an accepted ROLE UPGRADE, heading → body/eyebrow a 🟡 ROLE DOWNGRADE; any other role change, a CTA's included, stays 🔴 ROLE SWAP. Lost sr-only text is 🟡 MISSING SR-ONLY TEXT, except a link (a lost skip link stays 🔴 MISSING CTA). diff-profiles.mjs (both copies) gains the MISSING_SR_ONLY hint.
  • content-diff.mjs lists accepted upgrades on each root's block and in --json accepted, not as findings; header and usage name the new kinds.
  • Docs: diff/SKILL.md § Reading content-diff grades the outcomes; replica/reference/recreation-procedure.md § Role parity now says "mirror the look; semantics may improve" and points there, as do deploy/SKILL.md and replica/SKILL.md.
  • Tests: content-diff.test.mjs — each outcome, plus an end-to-end page: same-look <div> → <h3>, dropped sr-only text, and display:none / off-track-slide text (plain MISSING BODY).

0.34.9 — visual-diff flags what differs from the source, not what the build does alone (#142)

Field learning (a recorded replica gate on a home page): all 8 STRETCHED IMAGE flags at 1440 were stretched the same way on the source, burying the one real difference (thumbnails the source crops and the build distorts); heading faces could only be compared by hand.

  • diff/scripts/visual-diff.mjs: an image is stretched only under object-fit: fill. A build stretch is flagged only when its source twin — same rendered box ±2 px, same natural ratio ±4 %, never by file name — is not stretched the same way, and the flag names the source's fit ((source: object-fit cover at the same box)). The reverse — the source distorts an image the build crops — is an informational IMAGE FIT (info).
  • HEADING FACE (new): a matched heading at another size (> 1 px) or family. Families compare normalised (first in the stack, unquoted, lowercased, without Variable / VF / weight suffixes); one family pair across every differing heading is one HEADING FACE ×N line. Same-text headings pair in document order.
  • diff-profiles.mjs (both copies): the STRETCHED hint drops "unless justified"; new HEADING_FACE and IMAGE_FIT hints. diff/SKILL.md and the replica source-fidelity gate's pass bar name the two-sided rule and HEADING FACE.
  • Test (new): diff/scripts/test/visual-diff.test.mjs — the exported pure helpers (sourceTwin, normFamily, pairHeadings, headingFaceDiffers, redFlags), the ×N summary, --help without a browser, and an end-to-end source-crops / build-distorts fixture.

0.34.8 — the roster is reported against the site: a capped run never reads as the full site (#141)

Field learning (a recorded hands-off replica run): the setup status line said "full site" while the hands-off volume caps had cut the roster to 94 of 2,532 discovered URLs — 20 of 820 blog posts. The cap was recorded only as a named assumption in direction.md; the rollout ledger counted 94 as the total, and the owner learned the scope by asking.

  • stardust/SKILL.md § Hands-off defaults: "all pages" / --all-pages in the invocation lifts the caps; the setup status line names the roster against the site (roster N of M discovered — cap 100/20), and a capped roster is never called the full site. The locale bullet's anecdote folds.
  • rollout/scripts/inventory.mjs writes scope: { pages, discovered, capped, pageCap } into rollout.json from state.json's site.totalDiscovered / site.pageCap and prints a Scope line — "N of M discovered pages in scope — a SUBSET: the other K are not delivered by this rollout …" when the roster is smaller than the site. done-check.mjs repeats it under its verdict (and in --json when known): "complete" is about the roster. lib.mjs gains scopeOf / scopeLine.
  • Test: rollout/scripts/test/inventory.test.mjs — capped, whole-site and uncounted rosters.

0.34.7 — content-diff inventories page copy only: icon links, payloads and consent banners stop reading as missing content (#140)

Field learning (a recorded replica gate on a home page with no <main>): 12 of 53 structural 🔴 were MISSING CTA for menu links whose icon was <img> in the build and <svg> on the source; 9 🟡 MISSING BODY were inline payloads.

  • content-inventory.mjs (diff + deploy copies): a link with an icon and text is a text CTA, whatever draws the icon; only a text-less link is skipped. The label leaves out inline-svg <title> / <text>, and an image link's sr-only text is not a label. A card link — an <a> wrapping 2+ text blocks, or a heading beside an image — has its blocks inventoried as heading / body and is a CTA only for its own remaining text, so it pairs with an EDS card.
  • Not page copy: script, noscript, style, template and the common consent managers' banners (those CONSENT_CANDIDATES accepts) are left out of the role inventory, the Experience Workspace editable count and the attribute + icon layer; the root's summary line names a skipped banner (consent banner skipped (N text nodes)).
  • content-diff.mjs: icon anchors use the same label, so an svg-icon vs img-icon link is one 🟡 ICON KIND.
  • Docs: replica/reference/source-fidelity-gate.md rule 3 — --consent only adds an accept click; verify the banner is actually gone after dismissal.
  • Test: diff/scripts/test/content-diff.test.mjs end to end — svg- vs img-icon link, inline payloads, a consent banner with an icon close button, a card link vs an EDS card, an sr-only vs alt-only logo: no 🔴, the banner named in the summary. Every in-page consent list is pinned to CONSENT_CANDIDATES.

0.34.6 — the common consent managers are dismissed with no flag (#139)

Field learning (a recorded replica run on a site whose consent manager labels its button "Allow all"): no built-in candidate matched, so the bar was stitched into every live viewport chunk. The pixel round read 7.2 % at 1440 and 15.0 % at 360 (a FAIL) where the same page read 1.0 % / 1.6 % with the banner accepted, every hot band and section attribution pointed at the bar, and content-diff reported the banner's heading, link and buttons as 🔴 missing content.

  • diff/scripts/live-session.mjs CONSENT_CANDIDATES gains the accept-all controls of CookieHub, Cookiebot (dialog v2 and v1), Didomi, Usercentrics, Osano, Termly, TrustArc, Quantcast Choice, iubenda, Complianz and CookieYes, by vendor id or class, ahead of the text fallbacks, plus an "Allow all" fallback that must start the button text (a :has-text substring also hits "Disallow all") and sits before the bare "Accept" substring (which also hits "Accept only necessary"). --consent stays the override for a manager the list misses.
  • acceptConsent (new, live-session.mjs) is the one consent pass: each candidate resolves to its first visible match (CMPs repeat the accept class in a hidden preference panel), a click is credited only once that control is hidden, else the next candidate is tried, and one union probe makes a page with no consent control a single round trip. dismissOverlays (stitch-shot, the diff probes, chrome-parity) runs it, and so do extract/scripts/crawl.mjs and dynamics/scripts/lib.mjs settlePage when live-session.mjs resolves beside them, their own lists kept as the fallback.
  • Tests: test/live-session.test.mjs pins the vendor entries and the order (vendor selectors before the text fallbacks, "Allow all" before "Accept"); replica/scripts/test/stitch-shot.test.mjs captures fixture bars with no flag: an "Allow all" bar (consent dismissed via .ch2-allow-all-btn), a vendor class repeated in a hidden panel first, and an unlisted manager whose "Disallow all" and "Accept only necessary" precede "Allow all".

0.34.5 — replica: the source-fidelity gate can measure against a frozen origin baseline (opt-in, default off) (#138)

Field learning (a multi-day hands-off replica): the live site changed between rounds — a hero rotated, a banner appeared — and the gate's target moved under the fix rounds: a pixel FAIL measured the origin's movement, not the build, and each move burned a round. The gate captured the live side fresh whenever live.png was absent and had no way to hold the origin still. Nothing below changes a round unless asked for.

  • replica/scripts/stitch-shot.mjs --freeze-baseline [--force] freezes one origin per page and width: the capture always lands as baseline.png beside baseline.json { capturedAt, url, width, vh, height, settle, chunks, fontsFailed } in <out.png>'s dir, both written atomically; a freeze aimed at live.png is redirected there (one line) and never touches it; an existing baseline is refused without --force; a capture whose fonts failed is not frozen. --check-baseline captures a fresh shot and prints its Δheight vs the baseline — no verdict; <out.png> named baseline.png is refused before any capture. Two browserless subcommands carry the gate's bookkeeping: baseline <dir> [--url] [--width] (the capturedAt, exit 0 — or the one-line reason, exit 1: missing, a lone file, unreadable, frozen for another URL or width) and record <live.json> --origin live|baseline … (the round's live.json, atomic). playwright and pngjs load lazily, so --help and the subcommands answer without them. Tests: test/stitch-shot.test.mjs (new).
  • replica/scripts/gate.sh --gate-against-baseline makes the frozen baseline the origin side of the same pixel-compare: baseline.png is copied over live.png (sibling temp + rename), live.json records origin: "baseline" and originBaseline { capturedAt, height }, one stdout line names the capture time, no live shot. It holds the pixel target still only: a --full round's diff probes still navigate the live URL. A missing or stale baseline ends the round before any capture with one stderr line and exit 125 — never a fresh live shot in its place. Without the flag the round is the one it was: a live.png whose record says it is the frozen copy is retaken with one line (the record goes with it); a live.png with no record is a live capture, reused as before. A copy or record that fails removes live.png and ends the round (exit 1, one line), so baseline bytes never pass as a live shot. Tests: test/gate.test.mjs.
  • Docs. replica/reference/source-fidelity-gate.md § Iteration discipline states the rule once (its chrome-parity cost anecdote folded); the scripts index names the flags.
  • One record per artefact. A flagged round removes the capture record (live.capture.json, 0.34.4) beside the copied live.png, and the unflagged retake removes it with the copy; a --freeze-baseline capture writes baseline.json and no capture record. Merged with 0.34.4: one stitch-shot test file, both contracts.

0.34.4 — a 2xx denial page is a bot challenge: the capture refuses it, the gate round has no verdict (#137)

Field learning (a recorded hands-off replica run on a site behind a bot wall): the live origin answered the headless capture with status 200 and its denial text — about 900 px of "can't process your request … Incident Number" under one heading — and the gate accepted it as a normal capture: stitch-shot.mjs wrote the PNG and exited 0, no bot signal anywhere, and the pixel round measured the denial page as the source. The fail-loud rule (gate doc rule 12) read only the response headers, and a vendor wall that answers 200 leaves none to classify.

  • A 2xx denial page is a bot challenge (diff/scripts/live-session.mjs). After the settle gotoLive reads the page (pageDenial): a vendor signature — a phrase pair such as "Access Denied" + "Reference #" or "can't process your request" + "Incident Number" anywhere in the title and text; a single phrase ("Just a moment…", "Press & Hold", "Incapsula", "DataDome", "Request unsuccessful", "System Down") in the title, the text or an iframe src, only on a page under 2 KB of text — or a page under 2 KB of text under exactly one error heading. A match throws the same BotChallengeError the header path throws (marker HTTP <status> denial page — …; the message says "origin blocked" and names the --headed escalation); a full page that merely mentions a phrase is no denial page. Under --headed the solve window reloads up to three times before declaring it, and a reload the probe cannot read keeps the denial. Each probe read races a 10 s deadline; a probe that fails twice is one stderr line (denial-page check skipped), never a pass. Every instrument on gotoLive inherits the rule.
  • stitch-shot exits 3 on a denial page and records the check. The page is read once more after the dismissal sweep(s), before the capture (a denial page a script renders after domcontentloaded); a denial leaves no PNG and no record. A written capture now carries <stem>.capture.json beside it — { writtenBy, url, width, vh, height, chunks, denialCheck, at }, denialCheck = 'ok' or 'skipped: <reason>' — so a round whose check could not run says so in its evidence. playwright and pngjs load lazily (--help and the pure tests need neither).
  • gate.sh names the blocked round. A capture exit 3 ends the round with … capture BLOCKED (exit 3) — a bot challenge or 2xx denial page …: no verdict for this round; escalate --headed, never a pixel number; the --full probe verdict line says the same. Tests: diff/live-session (new — every signature, the 2 KB and one-heading rules, an ordinary page, gotoLive on stub pages), replica/stitch-shot (new — the re-check, the exit code, the record, --help; end to end where playwright is installed).
  • Live URLs only. The denial probe runs only where a vendor wall can exist (isLiveHttpUrl): a localhost or loopback build side serving a tiny page under an error heading is captured as-is; with a test. A single phrase in the title counts under the same 2 KB bound (an article titled "System Down" is no denial page). Merged with 0.34.3, so the two new test files are one per script (diff/live-session, replica/stitch-shot).

0.34.3 — consent overlays are dismissed on gate captures; the archetype ledger keeps its shape (#136)

Fixes from a recorded hands-off replica run on a site with a consent manager, verified against its artifacts.

  • replica/scripts/gate.sh dismisses overlays on its captures the way its probes do. The probes ran under --dismiss while the stitched live capture had no way to receive a site-specific selector: the run captured the live side with the consent banner and its scrim up, and the first pixel rounds measured them as page difference (18–50 % hot bands) until live.png was re-taken by hand. Both captures now run the shared dismissal by default, --no-dismiss turns it off everywhere in the round (new on stitch-shot.mjs as well), and --consent <sel> / --dismiss <sel,…> — stitch-shot's own vocabulary — reach both captures, the diff probes and chrome-parity; --no-dismiss with either is a usage error. diff/scripts/live-session.mjs knows the cmp-prefixed consent-manager buttons, so that manager needs no flag; its candidate list has a test. stitch-shot.mjs loads playwright and pngjs lazily and exports its pure parts (argument parsing, chunk targets, the scroll-stall guard, row stitching); test/stitch-shot.test.mjs covers them, --help without a browser, and a fixture capture where playwright is importable.
  • stardust/replica/progress.json is the map of every archetype — never one archetype's record. The residual-logging example in replica/reference/source-fidelity-gate.md showed a bare {pageType, archetype, breakpoints} object, and the run wrote exactly that as the whole file, destroying the types map, the siblings / migrate blocks gate-evidence.mjs writes and the _provenance block the resume and unit-boundary readers depend on. The example now shows the record inside types.<archetype> beside those blocks, with the rule that the file is extended per entry; gate-evidence.mjs wraps a ledger whose top level is one archetype's record under types.<archetype> (lossless, noted in its output) and refuses only a record that names no archetype; a ledger with a types map is never second-guessed; with a test; replica/SKILL.md names the keys.

0.34.2 — dynamics: a form's captcha is recorded and rendered, never silently dropped (#135)

Field learning (a newsletter sign-up on a media site): the live form carried a checkbox reCAPTCHA under the email field; the replica rendered the field and the button and nothing else. The form definition had no place for it, so the parity report read the form as done and the pixel diff saw a shorter card. A captcha is two halves: a public site key in the DOM, migratable, and a secret-backed server verify at the endpoint, which is not.

  • snapshot-forms.mjs: the definition gains protection — { vendor, siteKey, widget } from any [data-sitekey] element or a reCAPTCHA / hCaptcha / Turnstile script (v3 and Enterprise keys read from render=), null when absent; the per-URL stderr line flags PROTECTION: <vendor> next to REGULATED-PII.
  • forms.md § 4: a recorded protection renders the same widget in the same slot with the vendor's published always-pass test key and posts its response field with the data; the owner's real key, domain allowlist and server-side verify are one needs-backend decision row. Folded the "53 states / 260 countries" anecdote.

0.34.1 — deploy: CSS never lands in the document — a per-item look is block CSS or a named variant (#134)

Field learning (a Magento marketplace replica): the source home slider is PageBuilder, and each of its three slides carries its own overlay rule (padding: 200px 10px 10px 50px, … 750px, 190px 10px 0 35px; background-size: auto). The block ported that per-slide style as an authored config cell — padding: …; size: auto, documented in the JSDoc, tagged @ew-exempt config, copied into overlay.style.padding by decorate(). Two gates blessed it: the editability probe counted the three cells exempt ("24/27 editable, 3 exempt (config)" in the conversion log) and D15's lint knew only a full prop: value; } rule. The author saw CSS as body text in DA. Nothing in the skill said how a per-instance visual difference is encoded — only that it must not be flattened — and a page-builder source hands the agent a ready-made per-instance style field to port.

  • deploy ENCODE contract: new rule — per-instance visual variation is a NAMED VARIANT: every #90 fingerprint cluster becomes one plain word from a closed set of ≤3–4 (left | right, dark | light) in a dedicated cell on every row (or a block variant class); decorate() maps the word to a class, the block CSS carries each variant's exact source values, an empty/unknown word falls back to the declared default. That keeps fidelity (each slide looks as it did) AND authoring flexibility (an author can pick a side) — normalising clusters into one treatment is the exception, for sub-16 px deltas or when the set would exceed the D6/D9 budget, logged in the inconsistency register. Never a declaration, length, colour, class name or --token: value in a cell, and never el.style.x = text / setProperty(…, text) / setAttribute('style', text) from authored content in block JS. D15 names the bare-declaration form; EW5's "variant selectors" are keywords; anti-pattern 20; Step-8 JSDoc scaffold; checklist row.
  • @variants <cell>: a | b (default a) JSDoc tag: declares a variant cell's vocabulary. parseExemptTags() returns it; with a vocabulary declared, aggregate() reports an exempt ONE-WORD cell outside the set dead ("not a declared variant word") — a misspelled rigth would otherwise fall back silently in the block.
  • davids-model-lint.mjs: D15 gains a per-cell and per-paragraph check — text that IS a CSS declaration (every value token CSS-shaped: lengths, numbers, colours, functions, keywords) or a custom-property assignment is 🔴; a cell that is only CSS lengths or a lone hex colour is 🟡 (a data table may legitimately hold them). Key-value blocks (metadata, section-metadata, library-metadata) are exempt — configuration may hold a colour. Calibrated on 38k generated pages: zero 🔴 false positives ("color: text-only teaser boxes, 6-up grid" and decades like "1970s" were the two that had to be excluded), 204 hex-colour advisories on one site that had authored a brand colour per tile — the pattern this release names. The classifier lives in css-as-text.mjs.
  • ew-editability-probe.mjs (and block-roundtrip --ew, qa editability): an @ew-exempt tag no longer covers a cell whose text is CSS — aggregate() reports it dead with a #134 label, so the round-trip gate fails on the recorded case instead of listing it under "exempt (config)". Tests: deploy/davids-model-lint, deploy/ew-editability-probe (new).

0.34.0 — widgets: code-driven UI with no authored content ships as an EDS widget, not a block with empty rows (#133)

Platform change: the boilerplate gained widgets (adobe/aem-boilerplate #630, aem.live/docs/widgets) — a self-contained unit at widgets/<path>/<name>.html|.css|.js (+ optional .json) that an author mounts with one link (/widgets/<path>/<name>.html?<params>; params become data-* on the root), loaded by the stock blocks/widget/ and buildWidgetAutoBlocks(). Every skill modelled reusable UI as block, default content, chrome or module, and told the agent to build a search box, calculator or copy-free form as "one self-contained interactive block" with key-value config rows — the empty-placeholder block the platform now calls a widget. Four gates would have flagged a correct widget: rollout's .html link rule, deploy's round-trip harness (heading and CTA MISSING), qa's unblocked-page on a widget-only page, and the editability probe (dead text). The term "widget" also meant "bespoke content block" (a countdown) in deploy — a clash with the platform term.

One rule, stated once per skill: widget when the section carries no authored content — configuration only, or everything arrives from code, index, API or user input; block when authors fill rows. widget joins the reserved block names; a widget's file stem obeys the same -wrapper/-container guard.

  • Vocabulary (stardust/reference/data-attributes.md): data-widget="<path>/<name>" beside data-module; the mount is a link alone in its paragraph, never data-placeholder. DESIGN.json.extensions.widgets[] {id, path, params}; _meta.json.widgets[] {id, href} derived by migrate.mjs from a[data-widget], whose hrefs rewriteLinks now leaves verbatim.
  • deploy: Step 2 triage is three-way (default content / block / widget); #17 and #28 bound the interactive-block rule to sections that keep authored rows; scripts.js hooks name buildWidgetAutoBlocks() and blocks/widget/ as stock (ported verbatim on a pre-2026-06 clone); the fixed-asset grep scans widgets/; the bespoke-block sense of "widget" is renamed. block-roundtrip.mjs lists a /widgets/ link as widget:<name>, drops it before the role inventory and never inlines blocks/<name>/; ai-readability.md explains the default --exclude-blocks client-app,widget (the words come from code by design; strict still counts them).
  • dynamics: triage gains a fifth column, target (block · widget · runtime · content · embed); dynamics-plan.mjs emits it — widget for client-compute tools, the site-search box, form-less control groups, locators and kept chrome search/cart/account. A form with authored intro or success copy stays a block.
  • spec: kind gains widget on both sides; one CODED set in lib.mjs replaces the kind lists in spec-knowledge, spec-variants and spec-pick, so a widget is counted, variant-signed and capture-picked like any coded block; search is the starter widget (verdict: new); search, map, location-selector, chart and copy-free form families default to widget.
  • direct --prep / migrate / replica: a fourth prep action "mark as widget"; a widget instance renders as the mount link, so the placeholder gate does not fire; replica keeps "widgets are implemented, not justified away" and names the EDS delivery form.
  • rollout: ledger kind gains widget (blockPath = the code path, no document rows), pages schema gains widgets[], content-pending siblings inherit them, plan.mjs walks them; delivery-lint.mjs exempts /widgets/…html from html-extension and adds P1 unknown-widget (code file missing) and widget-name (reserved stem).
  • qa: a /widgets/ link counts as a block named by its file stem for missing-template-blocks and unblocked-page; new unknown-widget (HEAD on the code path ≠ 200); the shared editability probe keys the authored mount (<p> holding only the /widgets/ link) as widget and exempts it for both the deploy and qa callers. Tests: deploy/block-roundtrip, qa/widgets (new), dynamics-plan, migrate, spec-knowledge, rules, inventory, delivery-lint extended.

0.33.0 — martech: the source's Web SDK is read and offered as an aem-martech switch, off

Field learning: a full run on a consumer-brand site on Launch + Web SDK carried the tag stack over, but the hand-off only said "aem-martech is an owner project". It gave the owner nothing to act on. The Launch library already held everything the plugin needs: orgId, the Web SDK instance, edge CNAME, datastreams (two, chosen by a data element per brand domain), and a self-hostable 2.34.2 extension. It also showed 129 data-layer rules that only fire if the new pages push the same events. None of it was read.

  • spec-martech.mjs reads the Launch orgId and the Web SDK extension: version, self-hosted support, library type, and each instance's edge domain, defaultConsent and datastream ids (resolving a %data element% to the ids it holds). Rules now record their event types.
  • vendors.json gains a Web SDK row (adoberesources.net, alloy.js) carrying the aem-martech plugin facts: repo, path, default edge, the source's default consent, required Launch extensions.
  • martech-contract.json gains aemMartech, always enabled: false. Its config comes from that evidence. Ambiguous datastreams stay null, with the candidates listed. A rebuild keeps the owner's choices.
  • martech-handoff.md adds a data-layer section for any tag manager: the data-layer globals to keep, Launch rules with their push events and a count per event, the tag managers whose rules were not parsed, and a request for the owner's tagging plan. spec-martech.mjs saves every other tag-manager library (GTM and gtag by container id, read from inline snippets; Tealium, Ensighten, ... by src) to martech/tags/; the model reads them into the contract's tagNotes in Launch's shape (rules, data elements, consent groups, each citing from), which survive a rebuild and render in the same rewrite, data-layer and consent sections, marked as read by the model. The consent model comes from parsed evidence only; the model's reading is listed as evidence and never changes it. A Google Consent Mode default (inline gtag('consent', 'default', …), or a saved library that sets one) is saved to martech/consent-mode.json; inline on the pages, it makes the model google-consent-mode. A custom GTM data-layer name (l=) is kept. Its header adds dynamics-check.mjs after the scaffold re-run. It also adds a Web SDK section: the values found, the consent default to keep (in on the source, pending in the plugin), missing extensions, and the four switch steps.
  • martech-scaffold.mjs: with aemMartech.enabled, it loads the CMP, then initMartech/martechEager; martechLazy runs from loadLazy() after loadFooter, as the plugin documents, so page views do not wait for the delayed phase; martechDelayed loads the Launch property through launchUrls. A category starts the Web SDK pending whatever the source value and grants collect on consent. The scaffold refuses missing values, a missing plugins/martech, or the same property also enabled as a route. Off, the output is unchanged byte for byte.
  • Target is read the same way: the at.js extension (or a page-level at.js) and the Web SDK rules that render decisions. The contract gains target, off, and the hand-off a Personalization section. With target.enabled, the scaffold sets personalization, only on top of aemMartech. at.js activities stay an owner move.
  • dynamics-check treats the edge host as a route: forbidden on /, gated with a category, and expected on accept once enabled.

0.32.1 — spec knowledge, wave 4: what a reviewing client had to guess

Wave 4 of how clients use a spec's stardust/spec/knowledge/. A spec viewer that lets people review the spec inferred four facts with a model or recomputed them; the judgement now records them. All four fields are additive, so older judgement files still build (the fields read null).

  • open-questions.json default_option: the index of the option the default is, or null when the default mixes or conditions options. A reviewer sees the default as a selected option without matching prose to labels. On a sports-federation run, 23 of 28 defaults are one of their options. An index outside the options stops S10.
  • blocking is defined: confirm the default before launch, because a different answer means rework. The work proceeds on the default either way; read as "work waits", it contradicted "every open question ships a default".
  • findings.json chapter and rank: where a finding belongs in the spec's story (start, pages, build, dynamic, risks, decide) and its rank there, so a client places the agent's interpretation next to the numbers it explains and leads with the most important one. An unknown chapter stops S10.
  • redirects.jsonl needs_entry: 1 on a migration row whose path changes beyond the .html, index-name and trailing-slash rule. On the sports-federation run, 5,287 of 12,072 migration rows need their own entry. The template's redirect finding counted all of them; it now counts needs_entry=1.
  • templates/open-questions.json gives every generic question a default_option; templates/findings.json shows chapter and rank. Tests cover the four fields and the two stops.

0.32.0 — the header is a contract: its behaviour is recorded from the source and gated before done (#132)

Field learning: several runs shipped headers that passed every pixel and style gate while every menu was wrong. They had boilerplate one-level click dropdowns for hover mega menus, search and cart as plain links, and a static list for a drill-down drawer. Every visual instrument sees the header at rest, with the pointer parked and motion frozen. Deploy told the agent to keep the stock header's interaction code and treat the prototype's menu script as a visual reference. Nothing in the final verdict looked at the header.

  • diff/chrome-explore.mjs records the header by using it, at 1440 and 390. It hovers (desktop), clicks and opens from the keyboard every control, and records each opened state: links, headings, crop, aria-expanded, focus, scroll lock, motion, and close paths (Escape with focus return, outside click, toggle, pointer leave). It descends into tabs, flyouts and drawer levels within a budget, and a hit is reported. It types into search to record typeahead and the URL the submit lands on, and samples the header scrolled down, up and back. Navigations are answered 204, so the page never leaves.

  • diff/chrome-compare.mjs is the header-parity gate. It exits 2 on any of these:

    • a missing control or state (hover rebuilt as click included);
    • missing panel links or headings;
    • keyboard behaviour, aria-expanded, close paths or scroll lock that differ.

    The replica profile adds open-state crops (2 % bar, with the crop gate's texture so a font substitution is visible), motion (every duration within max(80 ms, 25 %)) and scroll states. functional, for redesign and reskin, checks behaviour and content.

  • Only the site owner clears a finding, in stardust/chrome/header-decisions.json; an entry signed by an agent is ignored and reported.

  • Wiring:

    • extract --prep records the contract (stardust/chrome/header-contract.json) and puts its summary in the prep summary;
    • deploy Step 6 builds the header from the contract, with the stock code as the accessibility floor and large panels as nav fragments;
    • deploy Step 10 and replica C0 gate the foundation on it;
    • done-check gains header_parity for migration projects (contract, a fresh comparison of the flow's profile at every contract width on the live host, no undecided error);
    • qa gains a chrome check;
    • dynamics-plan --chrome adds a row per control family (cart and account decide their backend, the UI ships);
    • spec S9 explores the home header and knowledge gains header.json, its size per width.
  • The replica eval gains header_contract_recorded.

Field run (a museum network's home page, a new EDS site):

  • The contract: at desktop, 6 click mega menus and a search overlay (14 states, 35 links); at mobile, a drawer with a search form and 6 accordion levels (14 states, 37 links). It took about 5 minutes for both widths.
  • The stock EDS header with the same nav content got 37 errors. Every accordion was flattened into one open list, search was a link, nothing had motion, the header was 64 px instead of 201 px, and the crops differed by 62–91 %. It correctly passed the desktop click, keyboard and close behaviour the source shares with the boilerplate.
  • A header block built from the contract reached 0 errors under functional. Under replica, 7 open-state crops remain at 2.1–3.9 %. They come from the licensed brand face replaced by a system font: an owner decision. done-check listed them as header_parity:7 and qa reported the same.
  • The field run fixed the explorer:
    • a <nav> carrying aria-expanded was taken for a control;
    • items scrolled out of a scroll container read as hidden;
    • icon words in aria-labels and transparent-font icons broke pairing;
    • POST search forms were not followed to their results URL;
    • crops took the links' parent instead of the panel box;
    • one duration per property lost a second animated element.

0.31.6 — spec knowledge, wave 3: what a migration still had to recover from raw files

Wave 3 of how other skills use a spec's stardust/spec/knowledge/. These are the facts consumers still could not get from the committed knowledge.

  • urls.jsonl blocked: why an error row was not read (challenge, refused, network, archive-miss). Gate 4 probes those instead of treating them as dead. On the museum field run: 61 challenge, 2 network, and no error row without a reason.

  • Search probes carry feature, path and param (the source search URL). dynamics-plan --knowledge drafts dynamic-features.generated-parity.json: one search-query check per probe, with the source's expectations and resultSelector left to the curator.

  • martech.json evidence: spec-martech's launch.json and onetrust.json, unchanged. martech.mjs readEvidence accepts a knowledge folder. On a sports-federation run (about 12,000 pages), the martech contract from knowledge is identical to the one from the raw folder: OneTrust, 4 routes, 58 rewrite rules.

  • archetypes.json (decision P3, step 2):

    • within each template, the layout variants covering 80% of its pages are page types;
    • the larger tail folds into the template's largest type;
    • a one-page variant or template is unique;
    • each type has one rep_url, so migrate's one-archetype-per-type rule holds;
    • extract --prep takes its types from it, and prototype --prep, replica and reskin take their representative.

    Every live page is placed exactly once: 29 types for 12,072 pages; 11 for 89 on the museum run, where the exhibitions listing became unique instead of borrowing the exhibition-page archetype.

  • Rollout pages record sourceUrl (from the sidecar, else state.json by slug; schema updated). qa from-knowledge and seed-redirects join coverage by it, so a renamed page still matches.

  • qa routing checks every expected redirect: a stardust/redirects.tsv row absent from the live /redirects.json is redirect-missing (error); a different destination is redirect-dest-differs (warn). Live, a planted missing row was reported, and the real 5 rows passed.

  • Additive fields: knowledge.md now states that a patch only adds fields and clients ignore the ones they do not know; removing or renaming a field is a minor release.

0.31.5 — spec knowledge seeds dynamics, qa, rollout and the archetype pickers

Wave 2b of how the other skills use a spec's stardust/spec/knowledge/: each consumer reads it when present and never requires it. Reading adds no wall time.

  • dynamics (dynamics-plan.mjs --knowledge [dir]): the spec's curated features become draft rows. They keep the axes as judged (S10 checks them against the dynamics vocabulary). Their decision is the linked open questions' answer, or else the default. It also works without a detector run. A decided-out row with no recorded owner answer is flagged (triage.md). Phase 1's search probes: a spec's search-probes.json already holds them.

  • qa (qa/scripts/from-knowledge.mjs, new): writes two inputs.

    • stardust/template-map.json: the inventory by layout variant at served URLs, delivered pages only when rollout's coverage exists.
    • Allowlist entries for links the source already had broken, as authored and as served.

    On a sports-federation run (about 12,000 pages): 114 variants and 1,251 source-inherited 404s that would otherwise read as migration defects.

  • rollout:

    • D2 starts helix-query.yaml from the spec's draft;
    • Gate 4 reads dead (http-*, redirect-broken, loop) and bodyless source pages from knowledge/urls.jsonl instead of probing them. An error row was never read (on the museum field run, 63 pages behind a bot wall), so it is still probed.
  • Archetypes: prototype --prep, replica Phase 3 and reskin default their representative page to the spec's rep_url of the type's largest layout variant.

  • Prose: rollout's query-index paragraph restated dynamics/reference/listings.md § Getting an index at all step by step; it now points there.

  • Field run (10 pages of a museum site on a new EDS site, with these scripts):

    • spec's draft helix-query.yaml, committed unchanged, indexed all 10 pages;
    • dynamics-plan --knowledge added 8 spec rows to 46 detector rows, all within the dynamics vocabulary;
    • qa's sweep with the derived template map ran on every delivered page.

    It also found three bugs, fixed here:

    • qa missed the 0.31.4 folder-index rule. It fetched /x and /x.plain.html, but a folder index lives at /x/ and /x/index.plain.html. The inventory now keeps each page's served url next to its key, and its slash probe checks that a folder index still answers at /x (the redirect spec records) instead of expecting a leaf's /x/y/, which 404s by design.
    • seed-redirects wrote redirects into pages of later waves. qa reported 5 redirects landing on 404s. With rollout's coverage, only redirects to delivered pages are written; re-run the seed after each wave.
    • Gate 4 would have dropped pages a bot wall hid. It counted outcome error as dead, and all 63 such rows on the museum site were pages behind the bot wall.

0.31.4 — a folder index lives at /x/: sitemap, probes and links follow EDS

EDS serves a folder index (x/index.html) at /x/, and /x 404s on a plain site (one site seen adds a 301); a leaf page lives at /x/y, and /x/y/ 404s (checked on a live origin). Stardust's rule was "never a trailing slash", which is right for leaves and wrong for folder indexes. Three consequences:

  • the rollout sitemap listed every folder index at a URL that 404s;

  • done-check's live probe (HEAD, no redirects, 200 only) counted every delivered folder index as not live;

  • localize-links wrote folder-index links that 404, and delivery-lint flagged the correct /x/ as a P1;

  • spec could not tell a folder index from its source path, so its redirects pointed at /x.

  • Rollout (lib.mjs publicUrl): pages[].path stays the key; a page whose migrated file is …/index.html is published, fetched and probed at path/ (assemble.mjs sitemap, verify.mjs, loadPageHTML, done-check).

  • localize-links writes the served URL (eds-path.mjs deliveredUrl): /x/ for a folder index, /x/y for a leaf; redirect destinations resolve the same way. Matching still uses the lookup key.

  • delivery-lint judges a trailing slash by its target, with the content tree (--content, default ./content):

    • a leaf's slash stays P1;
    • a folder index's slash passes;
    • a folder index without it is a new P1, folder-index-slash (it 404s);
    • a target outside the tree is a P2 advisory.

    First test file for the lint.

  • Docs: the rule ("no .html; a slash only on a folder index") replaces "EDS 404s on /x/" in deploy and rollout.

  • spec (servedUrls): a page with pages below it is a folder index, so its eds_path is /x/ and a migration redirect /x → /x/ is recorded (a plain EDS site 404s on /x).

  • Field rollout of 10 pages (4 folder indexes) on a new EDS site: every sitemap URL, done-check probe and localised link answers 200; /x without the slash 404s, as above.

0.31.3 — one delivered-path contract; spec redirects seed rollout

Five path rules disagreed on the same input: spec's edsPath, rollout's path-safety lint, deploy's canonicalPath and its copy in davids-model-lint, and rollout's own delivered path. /a/index.html gave /a/, /a/index.html and /a. Checked on live EDS sites: a folder index is served at /a/ (/a 404s on a plain site, /a/index too), a leaf page at /a/b (/a/b/ 404s), and .html 404s.

  • deploy/scripts/eds-path.mjs (new, pure): deliveredUrl (spec's rule, now also collapsing //), daPath (/a/ ↔ /a/index) and pathKey (the lookup key). spec's knowledge, localize-links and davids-model-lint import it, so the two canonicalPath copies are gone. spec's Setup copies the deploy scripts too.

  • rollout/scripts/seed-redirects.mjs (new): when a spec exists, Gate 3 first appends its redirects to stardust/redirects.tsv:

    • every source URL that changes on EDS;
    • the source's own redirects, pointed at the delivered page.

    Rows already there win; external targets, targets outside the knowledge and loops are counted and skipped. On a sports-federation run (about 12,000 pages) it seeds 12,312 rows, 241 of them existing redirects made one hop; on a museum run, 33.

  • Prose: rollout's path-safety rules are written once, in delivery-gates.md § Gate 3; two recorded anecdotes in deploy's link-localisation paragraph are folded.

  • A field rollout (10 pages of a museum site on a new EDS site) found two more:

    • localize-links rewrote links through any redirect, including the seeded ones to pages of later waves, so 8 links would have 404ed. A redirect now rewrites links only to a page of the content tree; the rest stay absolute.
    • The redirects sheet never matches a source containing : or % (those requests 404 first; measured with probe rows), and 31 of the museum site's 33 seeded rows had one. seed-redirects writes those to stardust/redirects-cdn.tsv, as CDN rules.
  • Open (next release): rollout publishes a folder index at /a, which 404s on a plain EDS site, and the link rules call a trailing slash a 404 although a folder index is served there.

0.31.2 — notes cleanup: two retired notes removed, dangling note citations dropped

notes/ held two documents nothing reads at run time, and six skill docs cited notes that were never in this repository.

  • Removed notes/deploy-improvements-archive.md (the frozen deploy ledger, findings #1–#80) and notes/multi-agent-distribution.md (the 2026-09 multi-harness assessment). The (#NN) citations stay as provenance; deploy/SKILL.md § References now says how to read the ledger from git (git show 940b8795:plugins/stardust/notes/deploy-improvements-archive.md). The README and replica/reference/handoff-contract.md no longer point at either file.
  • Dropped citations of missing notes: notes/variant-convergence.md (prototype/SKILL.md, the composition-delta-trivial fixture), notes/migrate-template-canon-refactor.md (prepare-migration/SKILL.md, stardust/reference/artifact-map.md, stardust/reference/data-attributes.md) and notes/prototype-broken-by-default-detector-2026-04-29.md (prototype/SKILL.md). The rules they backed are unchanged.

0.31.1 — spec knowledge, wave 1: complete redirects, the dynamics vocabulary, a scoping entry

A review of how the other skills can read stardust/spec/knowledge/ found gaps in spec's own output, and small readers that would fail on it.

  • Redirects (spec-knowledge.mjs): a migration row for every page whose delivered path differs from its source path. A path that changed only by its .html had no row, yet EDS answers 404 on .html and rollout wants a sheet row for each. On a sports-federation run (about 12,000 pages) the rows grow from 5,287 to 12,072.
  • Feature vocabulary: S10 checks class, disposition, reproducibility and status against the dynamics triage vocabulary (TAXONOMY, exported by dynamics/scripts/lib.mjs) and stops naming each value outside it. The field run's judgement used five values dynamics would have rejected later.
  • Router: a flow-neutral scoping entry ("scope / estimate / plan / spec the migration of X" → spec, never stamps flow). "Build a migration plan" no longer routes to replica. A large site (1,000+ sitemap URLs, no spec) gets a one-line offer to scope it first; it is never run unasked. The two copies of the wrong-flow anecdote are folded.
  • Ledger: spec's stages are phases (s1-inventory … s10-knowledge); spec writes its status lines.
  • localize-links --redirects reads JSON Lines ({src, target}, spec's knowledge/redirects.jsonl) and the published sheet ({data:[{Source, Destination}]}). It reported 0 pairs from the first and threw on the second. First test file for the script.
  • crawl.mjs --max 0 means no cap. It silently became 25, so a --all crawl translated to --max 0 kept 25 pages.

0.31.0 — publish always; done is one verdict

Two contracts a hands-off run had to be told from outside the skill, because the skill left them open. Both are now the skill's own.

  • Publish always (contract change). The replica handoff contract said "Publish is a decision… stay preview-only until it is made", while rollout Phase C says publish in the loop. Which rule won depended on which document the agent read last, and nobody owned the decision under hands-off. Now every delivered page, the redirects sheet and the chrome documents go live (PUT → preview → live). The decision step and its eds-conversion-log.md record are gone, C0 and the cluster deploys no longer hold --no-publish, and the DA protocol's publish step is the default rather than optional. deploy-batch.mjs --no-publish remains for an explicit preview-only ask. Deploy's font-licensing hold (self-hosted proprietary faces) is unchanged.

  • done-check.mjs (new, rollout). One completion verdict from the ledgers the run already writes:

    • pages_unfinished: coverage rows not verified.
    • pages_failed: failed rows, or a failed published-origin gate.
    • pages_not_live: delivered rows that do not answer 200 on the live host.
    • open_p1: the optimize gate's own count, source parity excluded.
    • rollout_incomplete: the last I-dashboard line is not an end.

    Exit 0 complete, 1 gaps, 2 usage. --offline skips the live probe, and --json prints { complete, gaps[], liveChecked }. Every reader that needs "is the run done" previously re-derived it from the same files.

  • Replica handoff writes replica handoff end only on done-check.mjs exit 0, and closes each gap with its own remedy (handoff contract § 5). A gap only the owner can close is a blocked line.

  • Bounded entry captures dynamics: extract --single --dynamics (and --pages … --dynamics), because Phase 2's dynamic-surface gate needs the per-page reach signals that --single alone does not record.

0.30.3 — spec: knowledge files instead of a database; one rule format instead of SQL

spec ended at spec.sqlite, but no stardust skill read it. Its one client, a viewer application, loads it into its own database, and the file carried that viewer's tables, onboarding text and HTML findings. The skill now ends at JSON files clients read, and knows no client. Contract change (released as a patch): knowledge/ replaces spec.sqlite, and judgement rules replace SQL.

  • Output (spec/reference/knowledge.md): stardust/spec/judgement/ (the agent's decisions) and stardust/spec/knowledge/ (computed by S10, spec-knowledge.mjs): urls, page blocks, blocks with variants, templates, variants, redirects, broken links, bad links grouped by target, features with their pages, martech, metadata, query indexes with a draft helix-query.yaml, locales, open questions, findings. Committed. Raw material, captures and crops stay in stardust/.work/spec/. On a sports-federation site (about 13,000 URLs) the folder is 37 MB, about 2.5 MB compressed.
  • Rule format (rules.mjs, knowledge.md § Rules) for feature reach, question impact and {{…}} in findings: URL terms (live, sitemap, block:, signal:, nested, verdict:), filters with globs, urls, count, sum, keyed and max: lookups. A rule that cannot be computed stops S10. sql:, url: and bvariant: are retired; block: no longer implies the sitemap (write sitemap block:x). No SQL engine is left in the skill.
  • Answers in judgement/answers.json replace a question's default in knowledge/open-questions.json; the default stays recorded.
  • No viewer concerns: no viewer tables, no tour or header label, findings as plain { title, text, numbers }.
  • Config at stardust/spec/spec.config.json; work names the raw-material folder.
  • Shared instruments: spec-martech classifies hosts with the dynamics skill's vendorFor (a missing table now fails instead of returning nothing); the index yaml follows the dynamics skeleton (description meta, image as a path, robots, the chrome and search excludes); Setup copies the diff and dynamics scripts, which S1, S2 and S9 need.
  • Verified on two recorded runs. Every rule from their judgement files (150) gives its SQL value on the same data, apart from an empty sum (0, not null). On the 13,000-URL run, spec-knowledge.mjs reproduces all 25 sets of its spec.sqlite, and a database rebuilt from knowledge/ matches it table for table (meta aside).
  • Field run on a museum site behind bot protection (generic profile, headed tier, 300-URL sample): S1–S10 ran in the new layout from the Setup copies, and a client database built from its knowledge. The CDN turned the headed browser away after about 100 pages; the spec covers the 89 pages read, and a finding says so. It found three 0.30.1 bugs, fixed here:
    • Drupal names: the generic namer took the first nameable class, so paragraph hid paragraph--type--<name> and every paragraph became one component. Block-style classes now win (lib.mjs, shared by parse and capture).
    • S4 refusals: a bot wall's 403 on a link counted as a broken link. spec-links records it as blocked, not a 404 (blockedBy moved to lib.mjs).
    • S1 crawl: the crawl kept <link> icons, fonts and manifests as pages (NOT_PAGE widened).
  • Existing projects: move <dir>/{inventory,fetch,parse,links,rum,martech,dynamics,media,sheets} to stardust/.work/spec/, judgement/page-blocks.jsonl and variants.json to .work/spec/map/, the config into stardust/spec/, then rewrite sql: rules and HTML findings.

0.30.2 — opt-in martech: carry the source consent and tag managers over, switched off

A migration dropped the source's CMP and tag managers silently, or left an agent to re-wire them by hand against accounts it does not own. Martech now has one contract, built only when the owner asks, from the evidence spec-martech.mjs already reads (dynamics/reference/martech.md).

  • Evidence: spec-martech.mjs --urls … --out stardust/martech runs standalone, without a spec.
  • Contract (dynamics-plan.mjs --martech): stardust/martech-contract.json and martech-handoff.md. It holds the CMP with its id, the source consent model (per-tag or owner-decision), the categories from the CMP configuration, one route per tag-manager loader, and the Launch rules and data elements to rewrite. Every part is written disabled, and a rebuild keeps the owner's choices. CMP scripts and loader patterns live in vendors.json; detection is unchanged.
  • Runtime (deploy/scripts/martech-scaffold.mjs): scripts/martech.js from the enabled parts. The CMP loads in loadEager(), tag managers from delayed.js, and a route with a category waits for that consent group. Tags load only on production hosts; on preview ?martech=on (plus &consent=accept) turns them on.
  • Verification (dynamics-check.mjs, qa dynamics): a martech check aborts cross-site requests after recording the host. It asserts nothing loads by default, gated routes wait for consent, and enabled hosts are requested after accept.

0.30.1 — spec: any website, a page cap, and blocked origins

spec parsed AEM sites only. A WordPress news site (2,061 sitemaps, about 412,000 URLs) and a Next.js site with utility-class markup showed what a generic parser and a cap need.

  • generic parser profile (spec/reference/config.md): below main, single-child wrappers are skipped and the children are components, named from a component attribute, a block-style class (wp-block-*, elementor-widget-*, paragraph--type--*), a BEM or CSS-module block, or a shape label (section.h2.list) where classes are only utilities; hash prefixes are dropped (stripPrefix for a site's own). One rule set (lib.mjs genericRules) runs in the parser and in the capture tagger: on 5 of 6 sampled pages the two found identical path ids. The AEM profiles are unchanged: a re-parse of 12,345 AEM Core pages was byte-identical.
  • One rule set per profile, one layout walk (lib.mjs profileRules, componentLayout): the capture tagger ran a hand-written copy of the AEM rules; parse and capture now share them, so path ids cannot drift. Parse output byte-identical on 12,345 AEM Core pages (5,192 grid rows), 200 AEM classic pages (651 column controls) and four generic sites; capture path ids and rows identical on the sampled pages of each profile.
  • Main region fallback for sites without main: [role=main], #main, #content, #main-content, article, then the body without header, footer and nav.
  • Inventory: stops when the origin redirects (two of three sampled sites were configured with the wrong host); crawls links from the scope root when no sitemap URL is in scope; --max / maxPages keeps an even sample per section and the database says so (meta.sample_note, inventory_total).
  • Template: a path-only rule ({ "pathSegments": 1 }) names templates by section alone.
  • Blocked origins (a museum site behind a bot challenge): S1 and S2 classify challenges with the diff skill's live-session.mjs (the edge signatures replica's gate uses) and stop at 10% (exit 3) with the choice for the user — allow-list the crawler, --headed (the plugin's stealth real-Chrome tier: 5 of 6 sampled pages on the first pass, where bundled headless Chromium got 2), or --archive <date> (Internet Archive raw captures, recorded per page and in meta.evidence_note). fetch/technique.json records the tier so later stages start in it, as extract's fetchTechnique does. Pages not read plainly are captured from the fetched HTML with live assets; a page still showing a challenge is an error, never a screenshot. Parse and capture agreed on every path id of the captured pages.

0.30.0 — new skill spec: the migration spec before the migration, as one documented database

Scoping a migration was rebuilt by hand per engagement: a crawl, spreadsheets of templates and redirects, a block-reuse guess, a list of "things to ask". Two recorded runs made the method a skill. The first, on a medical-technology corporate site (about 2,900 sitemap URLs, AEM classic components), produced the method and the database contract. The second, on a sports-federation site (about 12,400 sitemap URLs, AEM Core Components), ran from the skill alone. It ended with 114 layout variants, 28 blocks (7 reuse, 5 variant, 16 new), 27 dynamic features and 28 open questions, and no component left unmapped.

  • Stages S1–S10 (spec/SKILL.md): inventory, resumable fetch (every redirect hop recorded), parse, link check, optional real-user data, capture, mapping, layout variants, implementation, package. S1–S6 and S8 measure; S7 and S9 are judgement files under judgement/ with a rationale per decision.
  • Two parser profiles (reference/config.md): aem-classic (c-* roots, colctrl rows) and aem-core (grid members, column widths → rows; the root is the first non-aem- class). Lazy images (data-cmp-src, data-src, picture) count as images.
  • Template rules cover body attribute, body class or meta, plus pathSegments for sites whose CMS has one template for every page: the second run grouped by section, then by layout variant.
  • Mapping rules engine (spec-map.mjs): direct, layout, styled sections, nesting containers, carousel slides, row heuristics, dropIfEmpty / whenKids. Re-run until nothing is unmapped. Measured against the first run's hand-built mapping, 26 of 29 blocks came out identical and the variant clustering was identical.
  • Martech from public artefacts (spec-martech.mjs): headers and CSP, vendor hosts via the dynamics vendor table, tag-manager rules with path conditions, DOM selectors and the hosts their custom code loads, data elements that read the DOM, consent geo rule sets and categories.
  • Open questions ship a default (templates/open-questions.json, 24 generic ones); findings carry {{SELECT …}} for every number, so no figure is typed.
  • One output, spec.sqlite (reference/database.md): neutral meta keys and tables, plus consumer-owned tables (recorded decisions, saved views) that a reload keeps. Showing it is a separate application's job: the skill ships no viewer, and a consumer reads the database without crawling or judging.
  • Limits recorded: bot protection answered 403 to the dynamics detector on 28 of 44 sample pages in the second run (logged as a site-config item, not retried around). Content behind a role or rendered after load can differ from the parsed HTML; the capture tagger maps paths at DOMContentLoaded and falls back to the current DOM. A redirected locale tree is measured by final URL, not by the sitemap path.

0.29.1 — replica --target-breakpoints: deliver on the platform's steps, map by intent, gate the shift (#131)

A recorded replica migration of a medical-technology site (about 105 pages) moved its breakpoints to the EDS steps 600 / 900 / 1200 after delivery. Reused boilerplate blocks already sat on those steps while recreated blocks copied the source's (the gate rewards matching them): about 20 distinct breakpoint values in one codebase, spotted at once in review. Moving about 250 rules was quick; the before/after re-gate (about 105 pages × 6 widths × 2 builds) took a night against a live CDN. At 900 the pass count fell from 35 to 2, yet every failure checked was one of two intended changes (desktop header at 900 instead of 1024, about 40 px lower content; listings in their desktop row layout). Optional and off by default: without the flag, output, gate widths and register are unchanged.

  • Input and state (replica SKILL.md § Inputs): --target-breakpoints <list> (mobile-first min-width steps; --breakpoints keeps meaning gate widths), persisted once as DESIGN.json extensions.breakpoints.target — the one permitted addition to the promoted spec. The content cap is untouched.
  • Map in Phase 2, before any CSS (preserve-direction.md § 4): source switch points from the live stylesheets, mapped by intent (layout tier, not nearest number) into stardust/replica/breakpoint-map.md; 1px pairs collapse, max-width becomes min-width. Phase 3 and C0 apply the map, CSS and JS alike, and never decide it. One register entry, evidenced by a source | before | after strip.
  • Gate (source-fidelity-gate.md § Target breakpoints): widths add each target step; inside a shifted band a failure is tier-shift (logged against the register entry, not iterated, outside the cap; an overrides.json row at the delivery gate) or a defect (still fails). A noise check covers widths outside every band.
  • Lint (breakpoint-lint.mjs, new): every @media width, matchMedia string and innerWidth comparison under blocks/ templates/ styles/ scripts/ must switch at a target step, all query forms normalised to their switch width; skipped when no target is set. C0 wires it into npm run lint (handoff contract § Boilerplate lint), so deploy and rollout run it before delivery. --inventory lists a source's switch points for the map.
  • crop-compare.mjs --strip <out.png> [--c <c.png>] writes the bands side by side — the register entry's image from the gate's own captures.
  • Wording: extract's content-cap line and recreation's cap rule no longer say the target keeps its own two breakpoints (true of a redesign only); deploy's --nav-height rule follows the target steps when set.

0.29.0 — the capture freeze is capture state, never the spec: media must play (#130)

A recorded replica run shipped "poster + paused video" and static first slides across a site while 26/26 dynamics parity rows passed. Four causes, all in the skills: the gate defined truth as what the frozen instrument sees (stitch-shot pauses every <video> at t=0 and clears timers), so autoplay loops were masked as "nondeterministic media, permanent residual"; the interaction-parity rule "implement only behaviours that fired" read the frozen sampler's silence as dead; the freeze written as a deferral became the delivery brief verbatim; and parity's video-plays verified presence ("video elements present and controllable"), as did qa's rendered sweep. Contract change (a mask is now a promise, video-plays passes on playback), so a minor bump.

  • Capture state is a promise (replica recreation-procedure.md § Carousels, § Interaction parity; source-fidelity-gate.md § Residual logging format): every freeze or mask the gate needs is a captureState[] entry whose restoreAtDelivery names the live behaviour as observed (autoplay muted loop, rotation interval, marquee speed; null only when live is static too). Timer-driven media IS a fired behaviour — "dead" means never fired on live, never "stopped by the freeze". "Poster + paused" is a capture artefact, not a deliverable.
  • The list travels with the brief (handoff-contract.md fan-out discipline + Phase E row; rollout SKILL.md Phase C step 1, Phase E render check): C-deliver's brief carries the archetype's restoreAtDelivery list verbatim and implements it before done; Phase E asserts an autoplay video ≥ ¼ visible is playing and a rotating carousel advances unpoked.
  • The observer no longer blinds itself: motion-observe.mjs samples a media block before any scroll or poke — per <video> autoplay/loop/muted, paused and currentTime twice 1.5 s apart (playing = advancing), plus class mutations that fire with nothing touched (timer rotation). motion-compare.mjs gains the media class: live playing → build must play (a paused/poster build is MISSING), unpoked live mutations → auto-advance MISSING when the build has none; older observations without media still load.
  • Parity means playback: dynamics-check.mjs video-plays scrolls the <video> to ≥ ¼ visible and requires !paused + advancing currentTime; playbackHost is now optional (native video) and reducedMotionPauses reloads under prefers-reduced-motion and requires paused. Pure judgeVideoPlayback is unit-tested; patterns.md § media-as-url and parity-report.md rule 1 say presence is not a pass.
  • qa video-not-playing (rendered, error): an autoplay <video> ≥ ¼ visible that is paused, not advancing, or never buffered (readyState < 2 — the content.da.live 401 class) with src, readyState, paused and Δt as evidence; video-probe-failed (info) when the probe errors. Fixture-tested (judgeVideos).

0.28.0 — a migration keeps its design unless the ask says to redesign it (#129)

A recorded hands-off run resolved a bare "migrate" to the redesign flow (brand-faithful mode) because the ask carried no keep-design phrase; the owner expected a like-for-like move and got the prototype's reinterpretation — a static poster for the hero motion, a rebuilt mega-menu, pages that differed wholesale from the source. The master skill's third routing bucket ("anything else → ask; hands-off → redesign") was the root cause. Contract change, so a minor bump. Touches 0.23.0 (flow keys), 0.19.1 (zero-movement hand-off).

  • Routing (master § Two migration flows): the redesign flow is selected only by an explicit redesign phrase ("redesign", "modernise", "refresh", "new look", …); nothing is inferred. Every other migration ask is replica with flowSource: "default", no question, interactive and hands-off alike; the first response carries replica's "Say switch to redesign now" line so the one decision stays visible.
  • Flow keys (state-machine.md): flowSource is user-phrase | default | question; hands-off-default is gone. prepare-migration Setup stamps redesign only on an explicit phrase and hands anything else to replica; migrate's flow guard reads the same default.
  • Eval routing-migration-flow: prompt 1 expects replica without a question and flowSource: "default".

0.27.2 — the favicon miss becomes a failing check: favicon-default (#128)

Several recent first passes shipped the aem-boilerplate icon. The rule was already written in three places (deploy Step 3 § Favicon, replica's C0 foundation unit, prepare-migration assets prep) and extract had captured the icon every time; what was missing was an instrument — the QA check HEADed /favicon.ico and the boilerplate's own file answered 200, so every sweep passed. Prose repeated a fourth time would not have changed that. Touches #110–#112 (deploy favicon WARN), 0.19.1 (favicon HEAD at the origin).

  • qa metadata: resolves the icon the home page links (any rel~=icon, default /favicon.ico), fetches the bytes and reports favicon-broken (error, was warn), favicon-default (error: sha256 equals the boilerplate's favicon.ico — the site's icon never shipped) or favicon-mismatch (warn: differs from stardust/current/assets/favicon.<ext> when that capture exists — a format conversion is legitimate). lib.mjs fetchUrl gains binary: true; test/favicon.test.mjs.
  • rollout foundation-first gate runs qa.mjs --checks metadata against the origin and names favicon-default as the first-pass miss; deploy Step 3 § Favicon step 4 points at the same check instead of a bare HEAD.

0.27.1 — site search is self, never decided-out by a hands-off assumption (#127)

A recorded hands-off run of a 100-page medical-device site had the query index live and the plan script drafting the header search as index-backed / self, then wrote "site search out of scope" into its own named assumptions, let the triage table cite that as the owner's decision (decided-out) and left the header form posting to the source origin's results page — the one query-index feature dropped while its three listing siblings shipped. The start prompt had named search as a dynamic surface to handle. Touches 0.20.0 (dynamics triage) and 0.25.1 § Index, sitemap, search.

  • A hands-off assumption is the run's, never the owner's. It may pick an interim tier; it cannot set decided-out, which takes an owner's statement or a hard blocker. Dynamics SKILL.md Phase 3 and § Hands-off resolutions, triage.md § Reproducibility, the hands-off table in the stardust skill.
  • Search over the site's own pages is self once the index exists and ships over the migrated corpus; a corpus smaller than the source's is a parity note, not a reason to exclude. The header form never keeps the source origin's results URL (triage.md rule 4, patterns.md § search-index-backed).
  • dynamics-plan.mjs no longer poses "results page scope" as the decision on the search row; it reads none (corpus = the migrated pages; second corpora stay out).
  • Folded: the search ranking summary repeated in dynamics SKILL.md Phase 4 and triage.md rule 9 now points at patterns.md.

0.27.0 — archetype-first delivery: deploy and gate the archetype at the origin before its siblings render (#126)

A recorded 96-page run migrated 78 siblings from archetypes whose EDS encoding had never been gated, then re-deployed whole families in a five-workstream fix program; its sibling pixel bar ran on local renders whose numbers understated on every page. The flow now proves the archetype where it counts before anything inherits from it, and gates siblings once, where it counts.

  • Order. C-deliver gains unit C-archetype, one per template, between the foundation and its cluster: convert the gated archetype (the encoder every sibling reuses) → PUT → preview → the full gate row on the preview URL (gate-all.mjs --only, both widths) → fix rounds, cap 3 → done; only then migrate.mjs render for the template's siblings. Handoff contract § 1 intro and § 3 row C, progress.json kind archetype, replica Phase 5, rollout Phase C, migrate Phase 2, fidelity-tiers.
  • Siblings gate once, at the published origin. § 1 row 5 (the local prototype-regime sibling pixel bar) is replaced by the gate row after PUT + preview; gatesPassed pixel gates come from the page's row in gates/all-<w>/summary.json. Row 2: the structural clone is the encoder's INPUT, never a gated prototype; where it flattens a composition the encoder reads the hydrated capture.
  • What a failing row means (gate doc): one block → fix the block, re-gate every sharing page; content in the capture but missing → fix the encoder, re-render; a composition the archetype cannot express → re-prototype (new archetype when two or more pages share it, else a standalone one). Cap 3 fix rounds per page at the published origin, then that decision — never a fourth round.
  • Not in this release: a scriptable per-template encoder in the plugin (today the deploy convert step, proven on the archetype, plays that role); when it lands, migrate emits content/<path>.html directly.

0.26.0 — published-origin gate hardened: element-level criteria beside the pixel number (#125)

A 96-page rollout of a pharmacy retailer gated every deployed page on stitched captures and pixel-compare; its coupon-listing page PASSED at 6.7 % / Δh 69 px while all 284 cards were broken — a fixed-height card with overflow: hidden, the body ~30 px low, the description clipped mid-glyph, the details link pushed under the primary button, the sort control on another default. The links were in the served DOM: pure CSS geometry, which pixelmatch underweights (16.5 % over the text boxes alone). "PASS" meant "right shapes at the right places", not "every element present and legible". Element presence and legibility are checkable — so the gate checks them. Cross-references #115 (chrome crop gate), #124 (content-cap row) and the run's feedback items A1–A3.

  • New replica/scripts/gate-all.mjs (D0) — the all-pages published-origin gate: per deployed page in state.json, stitched captures of both sides, pixel-compare, then the two DOM probes below in the same window-free real-Chrome tier; per-page dir + summary.{json,md}, runs/ for --only. Verdict = pixel % ≤ 10 AND |Δh| ≤ 5 % of origin height AND clipped ≤ 0 (+ documented allowance) AND content MISSING + HIDDEN links / headings = 0 (+ required repeated units within 4 px); the pixel-only verdict is recorded beside it — every run is calibration data. Sidecars, each entry documented: masks.json, overrides.json (shown beside the number, never replacing it), clip-allow.json, presence.json, units.json. Origin fallback live → previous → crawl shot (asymmetric, flagged); --eds-host gates a code branch; a union (white-padded) pixel metric was tried as the height guard and rejected (white gaps score as matches) — the guard is explicit.
  • New diff/scripts/clip-probe.mjs (D1) — text lines and controls against every overflow- clipping ancestor: TEXT CLIPPED / HIDDEN, CONTROL HIDDEN / CLIPPED counted (exit 2); line-clamp, "read more" collapsibles (--more-words), scrollable containers, horizontal cuts advisory; collapsed menus, sr-only, hidden subtrees, off-page boxes never reported. Emits the visible text line boxes. Recorded page: 379; the live origin and two known-good pages: 0.
  • New diff/scripts/content-presence.mjs (D2) — visible headings, links (visible text), buttons, images, text blocks per band (h1–h3 aligned by LCS) + control state (--count-words); MISSING / HIDDEN LINK ×n and HEADING 🔴, buttons and CONTROL STATE 🟠, MOVED / EXTRA / COUNT / HEADING AS TEXT 🟡; symmetric scope, chrome left to the crop gate unless --chrome, --variable regions as counts (HIDDEN still counts), origin HTTP ≥ 400 fails loud (exit 4). Recorded page: HIDDEN 284, CONTROL STATE 2.
  • New diff/scripts/unit-geometry.mjs (D3) + diff/scripts/measure-live.mjs (library) — the shared settle / session / shadow-DOM measurement lifted from the project measure scripts; per-element Δx / Δy / Δw / Δh of the first N repeated units relative to the unit, hidden where clipped, origin cached per slug. Recorded card: body +28…+47 px, details link +28 and hidden, badge 70×70 at (−10,−10) vs 64×64 at (−27,−30), card 260 → 238.
  • pixel-compare.mjs --text-boxes (D4, auxiliary); textBoxPct exported. (--pad, the rejected union metric, is not shipped.)
  • The pixel-table rule. Every crafted prototype and every deployed page is a row: gate-all --stage prototype --proto-base <url> writes gates/prototypes-<w>/summary.{json,md} at the end of Phase 4 (reusing each archetype's cached live.png), the default stage writes gates/all-<w>/; gate-evidence.mjs reads the tables as the source of record (--tables; a page without a row is OPEN: no table row, --check names a missing table), the eval requires the prototype table, rollout Phase H reports both totals. gate.sh --full runs clip-probe on the build side in every regime, content-presence in the published regime, unit-geometry when stardust/replica/units.json declares a repeated-unit family for the slug ({ family: { origin, build, n, required, pages[], templates[] } } — replica recreation-procedure § Repeated-unit families); element lines fail the round and outrank a pixel PASS in gate-evidence the way the overflow assert does.
  • Coverage carries the verdict; the roster run is a recorded unit (review notes on the PR). update-coverage.mjs --gate <summary.json> writes each row into delivery.gate and flips a failing page (no documented override) to failed; verify.mjs never marks a page verified while delivery.gate.pass is false — completion derives from one place. Handoff contract § 3: each cluster subagent runs gate-all --only over its own pages inside the fan-out; C-final's roster run is unit gate-all (recorded running → done, resumable at its boundary) followed by --gate. Setup probes the deps before installing them (a harness that resolves them leaves the code repo untouched).
  • Upstreamed feedback. A1 launchStealthHeaded is WINDOW-FREE by default (STARDUST_HEADED_WINDOW=1 opts in). The tier is the BEST AVAILABLE and REGISTERED: Chrome when installed, else bundled Chromium with a warning, and every probe records tier (chrome | chromium-fallback | chromium) in its evidence; gate-all's summary has a browser column and counts degraded pages. A2 Akamai's HTTP 400 escalation is a challenge marker. A3 setup steps install playwright pixelmatch pngjs cheerio as devDependencies, never --no-save. deploy-batch.mjs appends the admin x-error header to 4xx (the DA 200-images-per-document cap; protocol step 3a').
  • Docs — gate doc § The all-pages published-origin gate; replica Phase 5 + Setup; deploy Step 10 item 7; rollout Phase E; diff SKILL § The published-origin probes; scripts index. Tests for the four instruments (pure parts browser-free; clip-probe end-to-end on a fixture where playwright resolves). Calibration on the recorded roster: pixel-only 66 PASS → full 28 PASS.
  • Open — default tolerances (--clip-max 0, --unit-tol 4, 25 % counts), how projects declare repeated-unit block families, whether buttons / control state should block.

0.25.2 — content cap: the container sizing model is measured, persisted and gated at a derived wide width (#124)

A hands-off replica run delivered a site whose live pages centre their content in two nested caps — a 1920 px page shell and a 1600 px content article — as an edge-to-edge build: the token was declared (--container-max: 1920px) and never applied, main > .section > div { max-width: none }, and BOTH pixel gates passed (1440 ≈ 1.2 %, 360 within residuals) because 1440 is narrower than either cap — the defect was invisible by construction. Extract had recorded the shell in a bespoke field and missed the article; the optional ≥1920 box-map spot check (#116) never ran, and pinned at 1920 it could not have seen the shell: a probe AT a cap's width reads the cap as the viewport. The two pixel gates are unchanged (1440 and 360, no third pixel gate); one cheap geometry-only read at a DERIVED wide width is added and made mandatory. Cross-references #13 (qa-gate's hardcoded 1600 wide pass) and #116 (the fluid-vs-fixed spot check, now folded into the row).

  • New shared instrument replica/scripts/cap-probe.mjs — extract, replica and deploy all call the same script. Capture: render a page at 1440, then read the same boxes at a derived width (max(2560, largest cap × 1.25); re-read once wider when a cap sits at ≥ probe ÷ 1.25) and at 0.9 × that width — resizes, not navigations; a box is capped when the two wide reads agree (a 90 % container scales, a max-width does not) — and list every cap ORIGIN (a capped box not merely inheriting, padded or auto-margined inside a capped parent) with its px, how it is authored (max-width / fixed / other), DOM tier and KIND — shell (outermost cap holding ≥ 90 % of the content root's text), content (a further wrapper inside it), module (one top-level section); chrome caps are not modelled. contentMaxWidth = the innermost wrapper cap, else the module cap shared by half the capped sections, else null (fluid). --write-design merges extensions.breakpoints.{containerMaxWidth, probeWidth, shellMaxWidth, caps[], modules[], capRegister[]} into DESIGN.json; several archetypes aggregate, and a cap that differs between archetypes or sits on a single module lands in capRegister for the inconsistency register (a cap on the shell or on every archetype is design intent, whatever its value; source breakpoints are never inherited). Compare (<live> --against <build> [--design DESIGN.json] [--slug <s>]): rows by KIND — wrapper cap, one module i/n per section (advisory while the section counts differ), content cap when neither side has a wrapper — never by DOM depth; PASS = every live cap held within ±20 px, nothing capped only on the build; a ✗ prints the scaffold rule to change (main { max-width } for a wrapper cap, the section div for a module cap), never a pixel target. Evidence lands in stardust/replica/gates/<slug>-<probeWidth>/ cap-<label>.{json,txt} like the pixel rounds. Exit 0 / 2 FAIL / 1 / 3 bot challenge / 125 usage; --probe-width pins the width for evidence runs only. probePage and the pure functions are exported; test/cap-probe.test.mjs covers classification, aggregation, rows, args, the DESIGN.json merge and — where playwright resolves — an end-to-end fixture (nested 1920/1600/1180 page vs a fluid and a fixed build, plus the pinned-1920 blind spot). scripts-index.md row added.
  • Extract records the sizing MODEL, not a number. extract/SKILL.md Phase 3: run the probe once over the archetypes (--write-design stardust/current/DESIGN.json, one navigation each, after the census); the extensions note names breakpoints; brand-surface.md § Spacing marks containerMaxWidth as measured; token-contract.md § Sizing: the replica flow inherits containerMaxWidth and probeWidth as-is, the step-up rules are redesign-only.
  • Replica pass bar gains the mandatory content-cap row. SKILL.md Phase 4 command + pass-bar bullet; source-fidelity-gate.md § Pass bar item 6 (the ≥1920 § Wide-viewport fluid check is folded into it — same-KIND pairing, re-lift upstream, no pixel iteration); recreation-procedure.md § Lift the sizing MODEL reads the measured cap model before lifting, encodes a wrapper cap on main and a module cap on the section column, and never reproduces a source breakpoint. evals/replica-source-fidelity: both_breakpoints_gated also requires the row's evidence; task.md names skipping it as a failure mode.
  • Deploy. Step 3 Foundation: the section scaffold honours the captured cap at its tier (main { max-width: var(--max-width); margin: 0 auto } for a shell/content cap, the boilerplate section div for a module cap, with the full-bleed escape for sections modules[] lists fluid). qa-gate.mjs --design DESIGN.json (default ./DESIGN.json): the wide pass runs at DESIGN.json's probeWidth instead of a hardcoded 1600, asserts the OUTER cap through the shared probePage (FAIL when the content cap does not hold within ±20 px of containerMaxWidth, or when the build caps a fluid source), then the per-block --maxw warnings (default now containerMaxWidth); no DESIGN.json or no probe beside the script → one ⚠ and the old behaviour. Step 10 item 5 is the cap row on the deployed URL; Local QA paragraph, pitfall 13 and the checklist follow; handoff-contract.md § 4 row updated.
  • Evidence (anonymised, 2026-09-24). A commerce home page. Capture: shell 1920px max-width main > div.<cms-list> (tier 1, holds 98% of the text), content 1600px max-width … > article (tier 2), contentMaxWidth: 1600px shellMaxWidth: 1920px probeWidth: 2560 consistent: yes. Pinned --probe-width 1920: only shell 1600px — the 1920 shell read as the viewport and vanished, the reason the width is derived, never fixed. Compare live vs the deployed origin: ✗ wrapper cap: live 1600px (content, … > article) → build NONE: content runs 2560px wide at 2560 (module caps only: 620px ×2, 340px ×2, 1180px ×1, …) — fix the sizing rule, not pixels — deploy Step 3 scaffold: main { max-width: 1600px; margin: 0 auto } → cap-probe: FAIL at 2560 — 1 of 5 rows failed. After that rule in the project's styles.css, served locally by the dev server: ✓ wrapper cap: live 1600px = build 1600px (±20) → cap-probe: PASS at 2560 — 0 of 5 rows failed; qa-gate.mjs --design on the same render: ✓ content cap 1600px holds at 2560 (#124). Second validation, the inverse defect on another delivered replica (a financial-services home page, --main at the live section wrapper): live modules are fluid (90 % framework containers, contentMaxWidth fluid), the build caps every section at 1430 px — ✗ content cap: build 1430px (module caps only: 1430px ×8) but live modules are fluid — fix the sizing rule, not pixels: remove the module cap → FAIL at 2560. Skill prose is net smaller than before (the spot check section, duplicated port-probe and hardening rationale were folded).

0.25.1 — replica run follow-ups: gates that passed a broken site, index/sitemap/search verification, capture gaps, the thumbnail cap, token files, recorded migrate units

Fixes from two recorded hands-off replica runs of one 36-page source site — one at 0.22.1, one at 0.25.0 — verified against their artifacts. The shared documents carry the same rules: replica/reference/handoff-contract.md (§ 1 rows 4, 5, 9 and the acceptance paragraph; § 3 rows D, D2, E2; § 4 rows for measure.mjs, thumb.mjs, ledger.mjs, the uploader, the batch driver and the rollout writers; § 5 ledger rule and migrate units), stardust/reference/scripts-index.md usage lines, state-machine.md site.captureGaps, journal-format.md's heading rule and replica/SKILL.md's pass bar.

Gates that passed a broken site

  • replica/scripts/gate.sh, measure.mjs, gate-evidence.mjs — horizontal overflow is a hard assert, not a residual. measure.mjs prints a root line per width per side (scrollWidth beside scrollHeight and the viewport, ◄◄ OVERFLOW +<n>px; root/rootDeltas in the JSON); gate.sh runs it on the build side every round (no live hit) and fails the round on an overflow of more than GATE_OVERFLOW_TOLERANCE px (default 4 — scrollWidth and clientWidth are integers, a correct page whose widest box is 360.4 px reads +1, and the qa skill's rendered sweep fails above the same 4 px; a real overflow is 13 px and up) whatever the pixel number says; gate-evidence.mjs records FAIL: horizontal overflow over a PASS pixel line, withholds pixel-gate-<w>, shows /ovf+<px> on the row and overflowX in the ledger. A recorded hands-off run delivered two pages 373 and 400 px wide at a 360 viewport and passed them as "logged, not iterated further (3-iteration cap reached)"; source-fidelity-gate.md § Iteration discipline names the assert no cap waives.
  • diff/scripts/content-diff.mjs compares attributes and icons, not text nodes only — on the main root AND the chrome roots. A second in-page inventory (placeholder, aria-label, title; small images by file name, inline svg, icon-font glyphs by ::before/::after content, css icons, empty icon boxes) and a pure differ — MISSING/EXTRA , MISSING ICON, ICON DIFF, ICON KIND, EXTRA ICON, ICON MOVED. Severity: MISSING PLACEHOLDER / ARIA-LABEL, MISSING ICON and ICON DIFF are 🔴 on an interactive element (input, button, link or inside one) and 🟡 elsewhere; MISSING TITLE is 🟡 always (a title tooltip is not read by most assistive tech and is commonly dropped by design); ICON MOVED, ICON KIND and every EXTRA are 🟡. Roots: --main takes a comma-separated selector list; --chrome (default on) adds the header and footer roots — the first <header>/[role=banner] and <footer>/[role=contentinfo] outside the main root(s) — and --no-chrome disables; the report has one block per root, every finding line carries its root ([header]), --json gains findings[] (with root) and roots{}; the Findings: line gate.sh and gate-evidence read and the exit codes are unchanged; playwright loads lazily so --help and the unit test need no browser. Recorded first: every delivered page lacked the search input's localized placeholder, a locale root carried the wrong flag, social links were empty boxes — and the probe read 0 🔴. Then measured on a delivered 36-page replica (5 page pairs): real regressions found on 4 of 5 pages — share links without aria-labels and icons, dropped CTA aria-labels, a dropped localized search placeholder — and two false-red sources of the layer's own removed: anchor-only icon pairing (7 present icons read as missing on one page — the build hosts an FAQ page's accordion glyphs in text-less <button>s anchored button#1…#7 where the source anchors each to its question; leftover icons now pair by identity and document order, within the same nearest block first, as ICON MOVED 🟡) and chrome outside the default root (the two chrome cases the layer was written for were invisible unless the caller passed --main header). source-fidelity-gate.md and handoff-contract.md § 1 row 4 say content-diff covers main and chrome roots on every gate page — the recreate and C-deliver commands need no new flag.
  • replica/scripts/measure.mjs reports the rendition per <img> (naturalWidth × naturalHeight, current source file; img.naturalWidth/img.file deltas) and recreation-procedure.md § Image renditions per breakpoint records which rendition live selects at each gate width and authors the same srcset/sizes; a sub-3 px root-height delta with equal boxes is a rendition-ratio symptom, not a layout bug (recorded: 0.64 px per inline image summed to 2.2 px and cost iterations).
  • Box model is a lifted value. boxSizing joins measure.mjs's default props so --against names a content-box/border-box fork per box; recreation-procedure.md § Box model is a lifted value lifts it per container, declares it per block and scopes the boilerplate reset on a replica, never universal (recorded: a universal border-box reset over content-box source grids — chrome −9 px, a facts row −106 px at 360 — while every prototype gated ≤ 0.1 %).
  • replica/scripts/gate-evidence.mjs no longer demands media-reconcile on prototype files. --content <dir> (default content); without the page's delivered content file the gate reads n/a: no delivered content file yet (never OPEN) and leaves the sibling acceptance set and --check; it is required again once the file exists. source-fidelity-gate.md places media-reconcile in the delivery chain (row C). Recorded: every sibling sidecar carried media-reconcile: OPEN until C-deliver.
  • replica/scripts/gate.sh retries a capture that exits 1 once. Live capture, build capture and the overflow probe; two exit-1 attempts end the round with exit 1 "re-queue, not a verdict" (3/4/124 stay final; run-bg slots unchanged at 3). Recorded: about 8 of 62 sibling gate rounds under parallel Chromium load ended build capture failed (exit 1) and were read as verdicts.

Index, sitemap, search

  • rollout Phase D / assemble.mjs — the served sitemap is verified, chrome is noindex. The assembled site/sitemap.xml is the EXPECTED set only (stardust/rollout/ is in .hlxignore; the platform serves its own /sitemap.xml from its index of published documents). Chrome and fragment documents (/nav, /footer, per-locale nav-* / footer-*, locale shells that are not pages) carry a Robots | noindex metadata row at write time; new --verify-origin <live-origin> fetches the SERVED sitemap (a sitemap index followed), compares its <loc> paths with the coverage rows, prints served / assembled counts with every extra and missing path, records manifest.json.servedSitemap and exits 1 on a mismatch; the D-site end ledger detail names the SERVED count. helix-sitemap.yaml + helix-query.yaml exclude globs documented as the additive alternative. A recorded hands-off run reported "sitemap 36 urls" from the local file while the served one listed 58 (every page plus 22 chrome documents).
  • rollout Phase D2 / dynamics — a missing query index is a yaml file, not a service. A 404 on /query-index.json means no helix-query.yaml in the code branch: author it (skeleton in dynamics/reference/listings.md § Getting an index at all — og:title title, description, image, lastModified; chrome and search documents excluded), push, publish live, poll no more often than every 5 s for at most 10 minutes; no configuration-service write is involved (a 403 there with the migration token is expected); the sheet-backed interim index is the fallback only when the code branch is not writable (triage rule 8, patterns.md). A recorded hands-off run read the 403 as "no index can be configured" and built the interim index.
  • dynamics search — parity is count + titles, not presence. patterns.md § search-index-backed: title matches rank first, description and body text only while fewer than N title hits (N = the source's visible count, read during detect), dedupe by title + description, cap the dropdown at N. dynamics-check.mjs search-query gains expectCount / expectTitles (+ titleSelector, countTolerance) and compares result COUNT and the top-3 titles with the source's recorded values, failing on a count mismatch or duplicates — pure compareSearchResults() with a fixture test (dynamics/scripts/test/dynamics-check.test.mjs). A recorded hands-off run's typeahead returned 10 unbounded entries for a term the source answered with 3.
  • rollout update-coverage.mjs --new — pages built outside the migrated tree enter coverage, once. --new <slug> --path </path> --template <id> --origin <origin> [--title …] [--status …] adds a schema-shaped row under the coverage lock (the pages schema allows no origin property: the origin is recorded in source.migratedHtml as <origin>:<slug>), creates or extends the template row, is idempotent by slug and refuses captured slugs and taken paths; assemble.mjs, verify.mjs --all and optimize.mjs include the row. inventory.mjs keeps such rows on a re-run (the origin marker identifies them; status untouched, one Kept report line) until the migrated tree holds a file for the slug or a migrated page owns the path — the page is registered once, never re-registered after every inventory run (inventory.test.mjs). A recorded hands-off run left its D2-built search page out of all three.

Bookkeeping, capture, tooling

  • ledger.mjs start guard. An end now needs an open start for the same skill + phase — an earlier start with no later end, paired in canonical forms; missing → warning, and under --strict exit 2 with nothing written and one stderr line naming the missing start command. run-status.md § Rules: the start line is the FIRST command of a phase, before any script runs. Recorded: a hands-off run wrote a phase's start and end one second apart after 109 minutes of work, so its supervisor saw an idle run.
  • ledger.mjs journal check. On end, a stardust/journal.md beside the ledger with no ## heading naming the phase (case- and separator-insensitive) gets one warning — never an exit-code change; an absent journal is not checked. replica/SKILL.md bookkeeping: every phase, extract and preserve-direction included, opens with the start line and closes with end plus a journal section ## <Phase name> — <what happened> (<date>); journal-format.md reconciles its ## <ISO-8601 timestamp> — <one-line summary> heading with it — in a stardust phase the summary starts with the phase name, and ledger.mjs accepts either shape. Recorded: one run's journal began at Session 2; another's had one section for the whole run.
  • crawl.mjs locale-root capture gaps. After the crawl, every captured locale root (a one- or two-segment path of locale codes — /en, /fr-ca, /ca/en) has the same-origin links on that page under its own path listed and the uncaptured ones written to _crawl-log.json#captureGaps (roots[{root, slug, linked, uncaptured[]}] + a ready detail string, uncaptured first-level targets: ca/en 15, fr/fr 0) with one stderr line; extract --prep copies roots[] to state.json.site.captureGaps (documented in state-machine.md) and carries detail in its end ledger line. crawl.mjs is now importable (playwright imported inside main, real-path main guard) and has a fixture-HTML unit test. Recorded: two runs captured /ca/en while 15 /ca/en/* pages existed and learned it six hours later from the rollout link audit (15 × 404, then repointed because 15 > the 12-page threshold).
  • thumb.mjs --max-bytes. Default 150000: a thumbnail over the cap is re-encoded at a lower --max-height (the same bottom crop) until it fits; every line now ends with the final size, a capped one also notes re-encoded for --max-bytes <b>; a cap unmeetable at one row is written, named on stderr, exit 1. extract/SKILL.md Phase 2.5: a thumbnail enters the context only after thumb.mjs has written it (≤ 150 KB), or the vision check runs in a subagent returning one line per thumbnail. Recorded: eight thumbnails of 268–608 KB — 3.2 MB — read into one context.
  • thumb.mjs --max-bytes scales the width before it crops, bisects on measured size, keeps at least 60 % of the page. The first hands-off run with the cap cropped 25 of 27 thumbnails to a median 28 % of each page (minimum 9 %) — a hero strip the vision check read as the whole page — and landed at ~7.5 KB under a 120 KB cap: the crop was the first lever and the byte-ratio shrink loop overshot by an order of magnitude (PNG size is not linear in rows). Now a thumbnail over the cap is re-encoded narrower first (480 → 400 → 320 → 240 px, the whole page each time), then shorter by bisection on the measured size, never below --min-share (default 60 %); still over at the floor, the floor thumbnail is written, named on stderr, exit 1. New --offset <px> thumbnails the rows below a crop (<dst> gains -<offset>), the stdout line names the scale and the kept share, and extract/SKILL.md Phase 2.5 has the agent read a cropped page's --offset slice too. pngjs also resolves through NODE_PATH.
  • da-media-upload.mjs and deploy-batch.mjs — --token-env <NAME> / --token-file <path>. The token comes from the variable named by --token-env (default DA_TOKEN) or from --token-file (read once at start, trimmed; never both; a missing or empty file is a usage error naming the path; the uploader's dry run reads no file); the expiry, HALT and re-run lines name the variable or file, never the value, and the file-mode re-run command carries no <var>=<fresh token> prefix. JWT-expiry preflight and the 401 policy unchanged; deploy/SKILL.md token paragraph: both drivers take both flags. Both test suites cover both flags (deploy-batch.test.mjs + 2 checks); the uploader drops an unused import. Recorded: the hands-off harness provided the token as a file while the uploader read DA_TOKEN only.
  • rollout optimize.mjs — findings that mirror the source are informational. With the extract capture present (--current <captureDir>, default stardust/current: pages/<slug>.json + rendered sidecar), a baseline finding whose condition the source shares — the same <title> (title-length / title-missing), no meta description on the source either, no JSON-LD on the source either, the same title or description shared by the same pages — is tagged fixability: out-of-scope with the evidence prefix source parity: and autofix unavailable (the findings schema has no parity property); listed in its own report section and excluded from the health score, the open P1/P2/P3 counts and the gate (lib.mjs: isSourceParity, markSourceParity, sourceParityCounts; computeScorecard skips them). A recorded hands-off run accepted 63 such findings by hand.
  • migrate renders as recorded units. The migrate phase — plan, the render of every sibling, state-and-report — is bookkept in stardust/migrate/progress.json with the C-deliver unit ledger's shape (units.<name>: {status: pending|running|done|failed, kind: plan|render|assets|report, templates[], pages: n, startedAt, endedAt, verdict}): one plan unit, one render unit per template cluster of at most ~8 siblings (migrate.mjs render <slug…> takes the slug list), one assets unit when media is rehosted there, one report unit. Per unit: the progress write, then the checkpoint commit, then the next unit — a unit end is a safe resume point; a session resumed inside migrate runs only the units not done, and a running unit re-renders its slugs (the driver is idempotent). handoff-contract.md § 1 row 9 + § 5, migrate/SKILL.md Phase 2. Recorded: on the 0.25.0 run the whole phase — 31 siblings through five parallel builders — ran as ONE unit inside one session, the most expensive session of the run, with no resume point between the plan and the final report.

0.25.0 — hands-off cost and autonomy: background instruments, inspection helpers, --help everywhere, bookkeeping writers, the Phase 5 contract

Source: Karl Pauls' stardust/replica-cache-perf (13 commits, recorded hands-off replica runs of 2026-09-18 to 2026-09-22), re-sliced into thematic commits, rebased over 0.22.2–0.23.1 and hardened after a code review and a cross-check against the stardust/next research branch. The theme is one number: a hands-off session's cost is its context, re-read on every model call and re-written whenever a step outlives the five-minute prompt cache.

Background instruments and the prompt-cache window

  • replica/scripts/run-bg.mjs (new): start detaches an instrument under run-capped (900 s default) behind a first-come slot cap; wait returns within --max (100 s, ceiling 110 s — inside the shell tool's default timeout of about two minutes, which applies only to a call that declares none; a declared longer timeout is honoured, and a recorded run completed waits of 179–181 s under a declared 200 s) with one line per job plus its verdict lines only, exit 75 while jobs are still going; log --grep reads the rest from stardust/.work/replica/bg/<job>.log; wait/status with no names report the latest batch. Recorded: one gate batch — three parallel steps, each sleep 5|10|15; then two rounds — blocked for 15 minutes and the next call re-wrote a 513k-token context, $6.30 for one silent gap, while every other gap in that session stayed under 184 s at a 96 % cache-hit ratio. Slots default to 3 (env override: RUN_BG_SLOTS, else STARDUST_BROWSER_SLOTS, lowers it on a small machine or raises it on a large one): each capture is a Chromium and a gate.sh --full round holds three at its peak, yet at 3 slots a recorded hands-off fan-out still queued 84 of 272 jobs for more than 10 s (90th percentile 78 s). Job liveness is checked against the wrapper's recorded process identity, never a bare PID (a recycled PID after a reboot read "running" forever and clean --all would have signalled a stranger); the wrapper is the only writer of its state file after spawn; clean --all escalates to SIGKILL on the process group.
  • gate.sh --full + --help: the whole Phase 4 probe set in one round — after the pixel verdict, content-diff, visual-diff and chrome-parity run in parallel under GATE_PROBE_TIMEOUT (300 s), reports to the gate dir, one verdict line each; exit 124 on any deadline, 2 on a pixel fail / structural 🔴 / chrome delta, 1 when a probe errored (no verdict is never a pass). When the pixel step itself gave no verdict (124, 1, 3, 4) the probes are skipped. The published round is the same command with the preview URL and --marker. The stale-instrument reaper matches node processes only, skips --inspect, and sends TERM before KILL.
  • Port probe without lsof (replica, deploy, diff cards; gate.sh): curl -sI localhost:<port>/ is the probe, a HEAD of your own file the reuse check; lsof … || echo free printed "free" beside a live listener on an image without lsof.
  • Cards: the master card's 0.22.2 wait discipline gains the harness-neutral form (run-bg, wait as the next call, never after a sleep, never two long instruments as parallel tool calls, the main agent bound like its subagents — a 338 s foreground turn cost one cache re-write); one shell one working directory; author large files in parts (~150 lines a write — two whole-document writes each re-wrote a ~200k-token context).

Inspection helpers and the reading discipline

  • replica/scripts/section.mjs, css-rules.mjs, json-query.mjs, html-slice.mjs (new): a Markdown outline or one section (multi-section --all capped at 20 KB); the rule blocks matching a selector regex with their @media condition; the shape, key union or a bounded filtered table of any capture JSON, --tsv / --path for values a command consumes; one element of a captured page, attributes stripped, scripts and inline SVG removed. Recorded: 172 steps carried 750k characters of tool output into a 535k-token context in 77 minutes.
  • extract/scripts/thumb.mjs (new): box-filtered whole-page thumbnails (--width 480) so 1-px rules survive the brand-gestalt read; the one exception to "one crop per fact".
  • replica/reference/reading-discipline.md (new; the replica card goes from 35 KB on the source branch to 31 KB): the rules in full — which helper reads what, instrument output only through run-bg.mjs wait and log --grep, one crop per fact, never a stitched page or the live/build/diff triplet, script flags from the index then --help, never the source. The replica card carries a 15-line pointer.
  • stardust/reference/scripts-index.md (new) + evals/lint/scripts-index.mjs: every shipped script on one line, kept complete by the lint (cap 12 KB).

--help on every script

  • Every script under skills/*/scripts prints its usage header on --help / -h and exits 0 before parsing, reading a file or opening a browser (recorded: section-schema navigated to --help as a URL, style-fingerprint crashed, davids-model-lint and sanitise read it as a file). Scripts that write artifacts carry a Writes: block. The six scripts whose playwright import preceded the guard import it lazily, so --help answers without playwright.
  • evals/lint/script-help.mjs (new, chained in lint:stardust and the validate workflow): runs every script with --help in an empty cwd; fails on a non-zero exit, missing usage text or a file written; pure libraries exempt by list.
  • Main-module guards by real path: node resolves the entry's symlinks for import.meta.url but not for process.argv[1]; a symlinked checkout made each CLI a silent no-op.
  • Value flags never swallow the next flag: --to --force, --name --dir and the like are a usage error naming the flag in the new parsers — run-bg, ledger, state, migrate, da-media-upload, the inspection helpers, thumb, measure, style-census, motion-compare (a status of "--force" would otherwise land in state.json).
  • impeccable-version-check.mjs --where prints impeccable's skill directory offline.

Bookkeeping writers

  • stardust/scripts/ledger.mjs and state.mjs (new): one conformant status.jsonl line per phase transition, the phase name normalised against each skill's table (--strict refuses an unknown one; optional --next / --owner); state.mjs advance <slug…> --to <status> follows state-machine.md (backward refused without --force, --history-only for the re-prototype case, the 0.23.0 flow keys kept in place). Recorded: one run wrote five archetypes' ledger lines a session later from a throwaway script.
  • migrate/scripts/migrate.mjs (new): the per-page render driver — page map, URL-literal output path (collisions checked on the AEM-folded path too), internal-link rewrite, asset bundling, <head> composition, strict validation, _meta.json sidecar, idempotent skip, the migrate block of state.json merged not replaced; gate / deviation / decision / variant / modules for the judgments only the agent records. Refuses to render on a project whose state.json carries no flow (0.23.0 routing).
  • replica/scripts/gate-evidence.mjs (new): derives the sibling acceptance gates from run-bg job evidence into gatesPassed[] and progress.json, attributing jobs by their arguments and taking the latest ended job per page × instrument × width. A sibling absent from a variance probe is OPEN, never passed on the archetype's summary; a job that hit its deadline or gave no verdict is OPEN; every pixel line names its regime ([prototype regime] on a local build URL, [published regime] otherwise); --check fails on a gate whose latest evidence regressed; writes are atomic. content-fidelity stays the agent's declaration.
  • replica/scripts/foundation-freeze.mjs (new): freeze writes the sha256 manifest of the delivery foundation (styles/, fonts/, head.html, scripts/, header and footer blocks); check exits 1 on any drift. Symlinks recorded, not followed.
  • rollout/scripts/inventory.mjs: an archetype whose sidecar leaves template null groups under its own slug; its representative is the archetype; empty modules[] counted.
  • rollout/scripts/lib.mjs blockCounts + update-coverage.mjs: a module the agent maps to EDS default content (title, text, image, button — no block needed) is recorded --block <id> --status converted --eds-name default-content and is never counted pending, whatever its status; blocks.mjs re-runs keep the mapping and the dashboard snapshot uses the same count. A recorded hands-off run left four such rows at status: pending, rollout.json.lastRun.blocks.pending stayed 4 with every page live and verified, and a downstream completeness check read the finished run as unfinished.

Media and delivery

  • deploy/scripts/da-media-upload.mjs (new): the DA media protocol as a script — multipart PUT, persistent ledger (atomic), capped retries, content URLs printed, the token never; --scope media/<name> refused (a recorded run doubled the folder, 242 wrong uploads). 401 policy from both recorded failures: a JWT past its exp exits 3 before any request; the first upload runs alone to prove the token and a burst 401 on it is retried (a valid token was answered 401 across a whole first burst); a 401 that persists after the token was accepted HALTS the batch (exit 3, ledger persisted, the same command resumes) — never a per-file FAIL (an expired token once left an 894-row batch failing file by file for 5.5 hours).
  • deploy/scripts/deploy-batch.mjs follows the same 401 policy: a JWT past its exp exits 3 before any request; a 401 on PUT, preview or live is retried like a 429; a 401 that persists through its retries on any page HALTS the batch — exit 3, no new page started, pages in flight finished, ledger persisted through the lock-safe merge, the halted page keeps its status (attempts incremented), one stderr HALT <path> 401 after <n> attempts — refresh DA_TOKEN and re-run: <the same command> line. Before, 401 was not retryable there, the page went put-fail and the batch ran on through every remaining page with exit 1 — the shape the uploader's rule was written against. Exit codes documented (0 / 1 failed pages / 2 usage / 3 token halt); --backoff-ms and the env test hooks DA_SOURCE_BASE, AEM_ADMIN_BASE, DEPLOY_VERIFY_ORIGIN (production defaults unchanged); contract test deploy/scripts/test/deploy-batch.test.mjs against a local fake of the three hosts.
  • rollout/scripts/media-reconcile.mjs: content-host URLs are decided hosted against the media ledger (--media-ledger, auto-detected), never fetched anonymously; a hosted URL missing from the ledger fails; with no ledger at all the URLs are unresolved (exit 1) — never a pass.
  • Shared ledgers safe under a fan-out (rollout/scripts/lib.mjs + update-coverage.mjs, deploy/scripts/file-lock.mjs + da-media-upload.mjs + deploy-batch.mjs): the coverage files, the media ledger and the batch ledger were read whole, mutated in memory and rewritten in place — with several cluster subagents recording at once the last writer won, and a lost media row then failed the media gate. Each writer now takes a cross-process lock (a lock directory, stale after 60 s, a 30 s bounded wait), re-reads the file on disk, merges its own rows over it and writes through a tmp + rename. Contract tests spawn eight concurrent coverage updates and a second uploader's rows.
  • deploy/scripts/build-harness.mjs: folds section-metadata blocks as the pipeline does and remaps DA image URLs to captured files via --media-ledger.
  • replica/scripts/measure.mjs and extract/scripts/style-census.mjs (new): box-by-box live-vs-prototype deltas for a selector list; the computed-style census over every captured page. Both open the live side through the diff skill's live-session.mjs (real-Chrome UA and headers, consent dismissal, bot challenge = exit 3), never a bare launch; the census runs once, in the background, after the crawl.
  • replica/scripts/motion-compare.mjs (new, advisory): one line per behaviour between two motion-observe outputs plus a summary; a MISSING / EXTRA line is confirmed on the class lines, never a gate by itself (the sampler misses class-toggled and pseudo-element mechanics).

Phase 5 contract and fan-out rules

  • replica/reference/handoff-contract.md (new): what the Phase 5 executor needs from migrate, deploy and rollout — sibling-tier steps, the editability, decode and DA protocols, rollout phases A–I with their ledger strings, one usage line per script, bookkeeping. C-deliver runs as recorded units in stardust/rollout/progress.json: C0 foundation authored and deployed by one foundation subagent, gated by the main agent on the published origin, then foundation-freeze.mjs freeze + commit; C1…Cn one subagent per template cluster runs the whole chain — local asserts, PUT → preview, the published-origin gates — driving its own deploy-batch.mjs --paths … --ledger stardust/deploy/ledger-<cluster>.json --concurrency 2 (the shared ledgers are lock-safe, so clusters deploy at once; a recorded 36-page hands-off run delivered 36/36 this way with zero 401s and its coordinator at or below 131k tokens of context) and reports one verdict line; the main agent only spawns, waits and records; C-final applies the queued foundation-requests.md once. A unit end is a resume point, never by itself a reason to end the session.
  • The foundation is a subagent's unit: a recorded run's main agent authored, deployed, gated and fixed the foundation itself in C0 and grew from 107k to 337k tokens of context inside that one unit — more than its four cluster units together; the coordinator now gates the shell only, and a failing gate goes back to the foundation subagent as one message carrying the numbers.
  • Rules placed where the agent acts (replica, rollout, deploy cards): lint and commit the foundation before any cluster subagent spawns, no frozen-file rename after fan-out; lint rules in every brief; disjoint clusters spawn concurrently; the root / is served by the index document and must answer 200 (a / redirect row only when the source root itself redirects); captured pages are always delivered — the ≤ 12 default is for uncaptured in-scope targets, which are a capture gap (crawl first) when the direction caps meant to include them and scope extension (repoint) otherwise, both recorded in direction.md; one-shot helpers live under stardust/.work/<skill>/probes/; every url() in a lifted CSS rule is rehosted under stardust/current/assets/.
  • Sibling pixel bar: each sibling is gated at 1440 AND 360 before delivery through run-bg (eleven siblings once failed the published 360 bar at 17–27 % after inheriting a gated archetype); recorded as prototype-regime evidence that never replaces the published-origin gate.
  • source-fidelity-gate.md: the header/footer crop-compare.mjs invocation is complete on the card, each band number named with the shipped output it comes from; the published round is the Phase 4 command with the preview URL.

0.23.1 — chrome: hover-dropdown reachability rule (deploy § 6, replica mechanism cloning) + qa dropdown-unreachable rendered check

Recurring chrome defect: a hover-opened desktop dropdown whose sub-list is absolutely positioned with an offset below the hovered <li> closes as the pointer crosses the gap; every crop and pixel gate passes because the resting state is identical (recorded 2026-09-19: 5/5 dropdowns unreachable on a deployed origin; an earlier same-symptom pointer-events: none case in the 2026-08 harvest).

  • deploy/SKILL.md § 6: "Hover dropdowns need a contiguous hover surface" — <li> spans the nav row with the sub-list at top: 100%, or keep the lifted geometry and bridge the gap with an invisible ::before (off in the mobile query; preferred in replica mode); pre-deploy grep.
  • replica/reference/recreation-procedure.md Mechanism cloning rule item 4: hover-path reachability is part of the state machine — verify trigger → first sub-link keeps the menu open.
  • qa dropdown-unreachable (rendered, desktop 1440, error): submenus discovered generically as header elements that become visible on hover over a header nav li; the pointer walks in 2 px steps straight down to the first sub-link; fails if the menu closes on the way or on arrival or the sub-link is not under the pointer. Once per distinct header; fixture test qa/scripts/test/dropdown-unreachable.test.mjs.

0.23.0 — routing enforcement: the migration flow is chosen once, recorded, and guarded at every entry

Evidence base: the same 48 field sessions. 0.18.5 fixed the routing surface (the two-flow table, "never mix", the never-chain descriptions) and the class survived it: on 0.21.1 a keep-design ask ("migrate … to the final fidelity") loaded prepare-migration; on 0.18.5 a "same design" ask adopted a redesign-only "train the template, then compile" plan and reverted it after an hour, 45 turns and 36 M tokens; a 144-hour resume session invoked no stardust skill at all and followed memory. Earlier in the corpus (0.18.1–0.18.2): "build a 1:1 migration plan" ran the redesign cascade for two hours before direct was asked for an "exact replica" and kept going — 2,207 pages published at 24–28 % pixel diff; a migration assessment ran the prep cascade 3.5 hours before the user asked for the keep-design flow and the work was discarded; migrate <url> / extract <url> as first commands on two same-design migrations led to hand-built compilers tuned by eye. Routing must be enforced at the sub-skill entry points and on the resume path, not only described in the master skill.

  • state.json.flow ("redesign" | "replica" | "reskin", + flowChosenAt, flowSource: user-phrase | question | hands-off-default) — reference/state-machine.md § Flow keys: stamped once by whichever entry resolves the choice (master routing, replica / prepare-migration / reskin Setup, direct's hand-off); changed only by an explicit --switch-flow, which records a MODE SWITCH in direction.md and marks the old flow's prototyped/migrated pages stale.
  • Master skill § Two migration flows: keep-design phrases ("exact replica", "1:1", "pixel-perfect", "faithful", "same design", "keep the current design", "re-platform only", every axis pinned) select replica without a question; redesign phrases select the redesign flow; anything else asks the one keep-vs-redesign question (hands-off: keep-design phrase → replica, else redesign, recorded). Planning aids are named per flow; a redesign procedure inside a replica run (or the reverse) is a routing defect to refuse or flag. § Routing: a resume (new session, "continue", memory-driven) starts with the state report and enters the next phase through its skill. § Journal rule: agent-authored crawlers, compilers, importers and gates that replace a skill phase are named deviations in direction.md. Description now says "resume a … migration".
  • Entry guards: migrate, deploy, rollout and a migration-intent extract refuse to start a migration on a project with state.json and no flow (two-flow table, hand back); prepare-migration refuses under flow: replica and resolves an absent flow before running; replica refuses under flow: redesign and stamps replica when absent; reskin stamps reskin. A bare extract for redesign/audit/uplift and deploy on hand-authored prototypes (no state.json) are unaffected.
  • direct Phase 1: a zero-movement phrase (every axis pinned) is not a direction — write the hand-off note, stamp flow: replica, stop with "run replica"; hands-off does not skip this. intent-dimensions § 9 says the same in one line.
  • First gate carries the choice: prepare-migration Phase 1 gate prints Flow: redesign … switch to replica; replica Phase 1 surfaces Flow: replica … switch to redesign. The state report gains a Flow: line and replica-flow recommendations from progress.json (last gate numbers per archetype).
  • Gated-archetype precondition (flow: replica): migrate (sibling tier) and rollout Setup read stardust/replica/progress.json and block a page type whose archetype has no gate result at each breakpoint that is a pass or over-the-bar with every residual carrying a cause. Reuses the gate's pass/residual semantics; thresholds unchanged.
  • evals/routing-migration-flow/ (new): four phrasings; asserts the flow is named first, the one question for plain asks, flow recorded, replica invoked and prepare-migration never loaded for keep-design, migrate never first, no hand-built pipeline.

Deferred to the T13 progress-surface work: the enriched state report as a script (stardust status).

0.22.2 — wait discipline: the coordinator never parks the conversation past the prompt-cache window

Evidence base: the same 48 field sessions (Aug–Sep 2026, 24 projects), 20.7k main-session requests. 1,271 requests re-wrote most of their context (649 M of 713 M cache-creation tokens); 1,222 of them followed an idle gap of ≥ 5 minutes. Classified by what preceded each: 46 % were the agent's own blocking waits — a foreground sleep (median 480 s), a 5–10-minute foreground gate sweep, crawl or publish loop, or a blocking read of an agent's output with a 10-minute timeout — worth 325 M tokens; 33 % were the user typing after a pause; 11 % were background-task notifications that arrived after 5 minutes. The cliff is sharp: a foreground sleep of ≤ 240 s missed the cache in 10 % of cases, 271–300 s in 18 %, ≥ 301 s in 85–87 %. Nothing in the plugin mentioned the cache window; the existing rule forbade sleep N; kill around instruments only.

  • Master skill, hands-off defaults: a "Wait discipline" bullet — long steps run in the background with a progress/summary file; the coordinator does independent work or checks the file at most every 4 minutes; never a fixed sleep ≥ 5 minutes, never a blocking output read with a long timeout, never a foreground command expected to exceed ~2 minutes. Ending the turn is for the user's benefit, not the cache's (a notification after 5 minutes misses too), so it is reserved for waits over ~45 minutes or decisions the user should take. A Claude Code-marked note names run_in_background and the harness's promptCacheTtl: "1h" owner setting (2× write price instead of 1.25×; on these sessions it alone would have cut cache-side spend by ~29 %, the rule alone by ~18 %).
  • Mirrors: deploy § 7 (brief/reading discipline), source-fidelity-gate § Iteration discipline (gate rounds over several pages run in the background, not as foreground for loops — 58 % of the recorded 5–10-minute sweeps re-wrote the prefix), rollout execution model (the batch driver's log and ledger are the progress file).
  • No instrument, gate, threshold or verdict changes; the DA protocol's capped until … sleep 3 waits (≤ 3 min) stay as they are.

0.22.1 — replica instruments: the pixel-compare hang fixed, deadlines, live-side caches, a path lint

Evidence base: 48 field sessions (Aug–Sep 2026, 24 projects) plus the notes those runs wrote about themselves. pixel-compare.mjs was slow or timed out in 17 of 24 runs and on every plugin version since 0.18.1; seven projects' own notes name it, four describe it sitting at 0 % CPU after printing its verdict; agents answered with sleep 150; kill loops that cost one page 30 fixed minutes. Reproduced 2026-09-18: with stdout redirected (as gate.sh and every agent pipeline run it), process.exit() after the compare hung Node's platform shutdown (Environment::Exit → DisposePlatform → WorkerThreadsTaskRunner::Shutdown → uv_thread_join) in roughly 1 run in 4; nine such processes from earlier migrations were still alive on the test machine, some for six days.

  • pixel-compare.mjs: exits by draining (process.exitCode) instead of process.exit() — 0 hangs in 10 runs where the original hung; plus a supervised --timeout (default 120 s, exit 124 = no verdict, not a FAIL) because the compare is synchronous and cannot time itself out.
  • run-capped.mjs (new): the deadline wrapper — macOS ships no timeout; kills the whole process group (Chromium children included), passes exit codes through, 124 on the deadline.
  • gate.sh: every capture runs under run-capped (GATE_STITCH_TIMEOUT 300 s, GATE_COMPARE_TIMEOUT 120 s), a partial live.png is never left for reuse, and this user's replica instruments older than GATE_REAP_MIN (15) minutes are reaped before a round.
  • chrome-parity.mjs --live-cache, anchor.mjs --cache: the live side's measurement is probed once per breakpoint and reused while URL, width and selectors match — the same contract gate.sh already had for live.png (recorded: 5½-minute chrome-parity rounds ×3 ×4 archetypes). anchor.mjs also warns when --main matched a wrapper that still contains header/footer (the false-structural-red class recorded on four archetypes at once).
  • evals/lint/script-paths.mjs (new, in npm run lint:stardust and CI): every plugin-internal skills/<skill>/scripts/… and skills/<skill>/reference/… path a skill doc names must exist. It found two dangling references on main (a gate example pointing at the wrong skill's content-diff.mjs, a hypothetical migrator path); both fixed.
  • Docs: source-fidelity-gate § Iteration discipline names the deadlines and caches; replica Phase 4 snippet uses them; deploy § 7 and the master skill's hands-off defaults get a brief-size and read-by-section rule (across twelve runs deploy/SKILL.md was read whole ~20× per run, once per dispatched agent; in one recorded run the watchdog-killed agents carried the fattest briefs); the DA protocol's two until … sleep 3 waits are capped and fail loud.

0.22.0 — stardust owns stardust/: write boundary and versioning policy

Field projects answered "what do I commit" by hand, each differently; one lost its state machine to a bare state.json exclude. Now: stardust writes only to stardust/, the impeccable root files and the EDS project (via deploy / rollout / dynamics); run-only files go under stardust/.work/. Three leaks fixed (replica's script copies left the boilerplate scripts/, deploy's harness left root qa/, deploy's pre-render left samples/). New Setup step 6 installs stardust/.gitignore (everything tracked except screenshots, current/assets/, replica/gates/, migrated/assets/, rollout/qa/, .work/, run residue, session state), covers .env, lists stardust/ in .hlxignore, asserts state.json is not ignored, offers LFS above 50 MB. State report gains a Repo: block; artifact-map gains § Versioning; qa baselines are local.

0.21.1 — evals: criteria.json in the tessl weighted_checklist schema

tessl plugin publish validates every evals/*/criteria.json against the registry schema (context, type: "weighted_checklist", checklist[{ name, max_score, description }]); the plugin's evals used the runner's own { criteria[{ id, weight, description }], total } shape, so 0.20.0 and 0.21.0 both failed to publish. All eleven rubrics are converted (ids → name, weights → max_score); the runner normalises either shape (normalizeCriteria), so scoring is unchanged.

0.21.0 — AI readability: the checker formula, one gate, document-first listings

Four migrations (a family-entertainment chain, a semiconductor company's replica, a UK package-holiday retailer, a beverage brand pilot) were scored 40–58 % by Adobe's "AI Content Visibility Checker" while every stardust gate was green. Each session reverse-engineered a different model of the tool — served-text parity, hidden text, markdown line diff — and each spent a round on a fix the score did not reward (inlining nav/footer into 143 documents; clipping instead of hiding; unwrapping generated anchors). The extension's own analyzer code settles it: score = min(100, served words ÷ rendered-DOM words), landmarks stripped by default, hidden text counted as rendered, a count ratio and not a word-set diff. Reproduced to the word on one site.

  • New reference deploy/reference/ai-readability.md (on demand): the formula and its consequences (what JS adds to the DOM is the whole defect; hidden text, chrome and generated anchors are neutral; served-only text inflates; short pages suffer most), the two metrics kept apart (checker score vs served-text parity), a cause-class table with remediation (loop clones, index-fed cards, runtime fragments, definition-driven forms, generated labels), six block rules, chrome inlining as a documented option with its trade-off, and the gate contract.
  • New gate deploy/scripts/ai-readability.mjs: exact reimplementation, both toggles, code score (fragments credited, app blocks excluded), per-block served-gap attribution, allowlist by block + string, JSON report, exit 1 below --min (98). Runs in the deploy atomic delivery contract on the published page, as the new qa check ai-readability (K), and in audit Phase 4.
  • Block rule (deploy, always-on, one bullet): decorate() adds no words to the DOM — clones presentational, listings document-first, generated text only for allowlisted runtime values. The D12 key-facts paragraph shrinks to a one-liner that points at the reference (net always-on growth ≈ 0).
  • Listings contract rewritten (dynamics/reference/listings.md): document-first — one authored row per item with the card's text, a heading row per group, a label-list row; the block uses the index for non-text fields and top-up; re-runs replace their own rows. Dynamics Phase 4 and rollout D2 point at it.
  • Replica: loop clones carry no text/alt/href/aria (recreation-procedure.md).
  • Eval evals/ai-readability/: clones, document-first listing, explicit fragment decision, no generated text, gate reported, checker facts stated correctly.
  • Not adopted, on evidence: clip-instead-of-hide rules (hidden text is neutral), inlining nav/footer into every document as a default (does not move the default score; two of three owners declined), CSS-stretched links for the score (kept as the accessible-name shape only).

0.20.0 — dynamics: the dynamic surface of a migration

Three real migrations (a US health insurer's employers section rebuilt greenfield, a UK package-holiday site on an existing EDS library, a consumer-credit site re-platformed at ~5,900 URLs) found the same thing: a site's dynamic surface — modals, players, forms, search, tags, APIs, client-rendered and sheet-backed content — is invisible to a block-scoped, pixel-verified pipeline, and invisible in a way every gate certifies as correct. This release makes it visible, forces a decision per row before import, and proves the behaviour after delivery. Migration-bound by design: default-on in both migration flows, never for redesign-only work.

  • New sub-skill stardust:dynamics (skills/dynamics/): detect → classify → triage → implement → verify. Triage on four axes — class (L S F M V T A R X I18N CR D), disposition (rebuild-native · index-backed · data-fed · embed-passthrough · client-only · static-snapshot · decided-out), reproducibility (self · needs-credential · needs-human-capture · needs-backend · needs-business-decision), status. Only self ships autonomously; the rest is one owner decision batch. Hard rules: reconcile against the migrated output first, never fabricate a blank client-rendered page, never auto-wire a regulated-PII form, a search box implies a results page, decided-out is explicit.
  • References (loaded on demand, not in the always-on skills): classes-and-signals, triage (+ the stardust/dynamic-features.md inventory format, which subsumes the former dynamic-blocks map), patterns (catalogue with contracts and three embedded example mechanisms — modal loader, index search, JSON post with honeypot), listings (folded from rollout's dynamic-listings), off-origin-data (host-keyed endpoint indirection, code-bus snapshots, fetch shim, per-state rendered snapshots + Source row, sheet sync, chrome URL space), forms (controls not form tags; intake by content source; regulated data), parity-report, locale-trees. The plugin ships contracts and tooling, not blocks: no block was reused as-is across the three cases and an existing library must be fed, not forked.
  • Tooling (skills/dynamics/scripts/, verified end to end on a local fixture): dynamics-detect.mjs (network log by host, first-party API paths with status, POST bodies, forms and form-less control groups, the trigger → dialog → content graph, player ids, iframes without src, tag-manager mount divs, settings-object keys, framework, auth/commerce/locale, client-rendered slots and pages, listing candidates; --from-state for one page per type; --reach folds the crawl's per-page signals in), dynamics-plan.mjs (four-axis draft per finding, --target-origin host-bound probe of every recorded API path, --migrated reconcile against delivered output, regulated-PII flag), dynamics-check.mjs (parity replay over a closed set of check types — fetch-json, dom-count, click-dialog, search-query, form-flow, video-plays, consent-gate, no-page-errors — third-party request statuses recorded per check), snapshot-api.mjs, snapshot-forms.mjs, sync-sheets.mjs, vendors.json (the classification engine: host pattern → class + role, product names only).
  • Hooks in the existing skills, ≤15 lines each: extract --dynamics (opt-in per-page reach signals; never set by a bare extract, uplift or audit); prepare-migration Phase 4.5 and replica Phase 2 run Phases 1–3 as the pre-import gate; migrate Phase 1 is the safety net for the hand-run flow; deploy reads the inventory as brief input and never flattens client-only / modal-bearing sections into prose; rollout B2 verifies the inventory against fresh evidence, D2 implements the self set and batches the rest, H reports parity; qa gains the dynamics check (parity-missing / parity-failed / parity-env-limit / parity-unchecked); master routing + both flows name it.
  • Origin-scoped site auth everywhere. resolveSiteAuth / attachOriginAuth in the shared live-session helper and the qa runner (--auth-header / --token-env, default SITE_TOKEN): the secret rides a route filter on the base origin only — a context-wide header leaked it to a video vendor's playback API, whose CORS check then produced a player error real users never see.
  • Learnings ledger failure classes dynamic-gap / api-dependency now point at the dynamics references. Removed: rollout/reference/dynamic-listings.md and the 0.19.9 dynamic-capabilities.md (folded into skills/dynamics/reference/).

0.19.8 — impeccable dependency: unpinned by design, with an update hint

  • Dependency declaration moves to the documented cross-marketplace object form — { "name": "impeccable", "marketplace": "impeccable" } — still with NO version range, on purpose: impeccable's design craft should always be the current one. (The root marketplace's allowCrossMarketplaceDependenciesOn already lists impeccable.)
  • New stardust/scripts/impeccable-version-check.mjs — Claude Code only announces plugin updates through marketplace auto-update, which is off by default for third-party marketplaces, so a user can sit on an old impeccable indefinitely. Setup step 1 now runs this check once per session and surfaces its one line when a newer impeccable exists (installed version from the plugin registry or --local <dir>; latest from the upstream manifest with a 6s timeout, falling back to the cached marketplace catalog; --offline, --json). Advisory only: always exits 0, fails silently to "unknown" — the registry paths it reads are Claude Code implementation details, not an API.
  • Manifest drift fixed: the adobe-skills marketplace entry and the Tessl manifest both still said 0.18.1 while plugin.json was at 0.19.7 (the validator warned; plugin.json wins at install, so users were unaffected, but claude plugin tag requires agreement). All three now read 0.19.8.

0.19.7 — sibling variance probe (P12)

  • New replica/scripts/sibling-variance.mjs — before cloning a gated archetype onto its siblings, probe the template-defining computed values on every sibling's LIVE page and diff against the archetype: per --probe name=<sel> the match count, first match's box + computed group (background layers incl. gradient scrims, colour, padding, font, radius), first heading and image, list-style and ::before mechanism, and the number of distinct style families among matches; plus the top-level section list. Defaults (first section, most-repeated class, li) when no probes are given. Live-session hardening as the other replica instruments. Exit 0 constant, 2 variance found. Read-only — it never edits the clone. Field evidence: eight "same-template" siblings varied in hero template (441 vs 528px), scrim direction, bullet mechanism and terms shape — all found late at the pixel gate.
  • Fidelity tiers: the sibling tier is "variance-probed" first; new § Sibling variance probe — every delta is budgeted as a block VARIANT class emitted on the sibling's content (blocks stay generic, never forked per page); gatesPassed gains variance-probe, _meta.json gains variants[]. Replica Phase 5 and migrate's A′ branch point at it.

0.19.6 — replica: glyph-dense chrome noise floor, evidence-gated (P6)

  • crop-compare.mjs reports the diff TEXTURE — the share of differing pixels with ≥5 differing neighbours: thin-edge = glyph-antialiasing noise, thick = blocks/bands (misalignment, missing paint). Reported in text and --json; never changes the exit code.
  • Gate doc, pass bar item 5: a glyph-dense chrome band that fails the 2% bar may be logged as a justified residual (cause: "glyph-antialiasing") — never a pass — only when all three hold and are attached as artifacts: chrome-parity.mjs exit 0 for the region at 1px tolerance, crop-compare texture thin-edge (≤15% thick), and a text-dense region (no imagery/icons). Field evidence: a ~50-link footer bottomed out at ~5% with every metric numerically identical; a hinted licensed face vs a self-hosted webfont rasterise differently per glyph. The 2% bar is unchanged; residual logging format gains the entry shape.

0.19.5 — deploy: link localization as a pipeline stage (P24)

  • New deploy/scripts/localize-links.mjs — dependency-free, idempotent. Builds the URL map from the content tree (every served path, extensionless, x/index.html → /x) plus --redirects (rollout's stardust/redirects.tsv or a JSON map), rewrites every source-host <a href> (with/without www., http/https/protocol-relative) whose path resolves in the map to the canonical root-relative form — no .html, no trailing slash, query and fragment preserved — and normalizes root-relative internal hrefs the same way. Everything else stays absolute and is REPORTED (the not-yet-migrated boundary). --dry-run, --json, and --check (write nothing, exit 2 when localizable links remain — the pre-deploy assertion). Field evidence: ~500 source-domain links across ~150 pages bounced visitors back to the live site for pages that existed on the new origin.
  • Deploy stage: the Deploy table and the per-page atomic delivery contract gain the stage (run after every generator and before every write, over the WHOLE tree; re-run after every wave). The ENCODE D4 bullet and davids-model.md now say what D4 is — a capture-fidelity rule for media and external targets — and that internal links to migrated pages are root-relative. Checklist item added. Rollout Phase E2 points at the stage instead of a hand rewrite.
  • Lint (advisory): davids-model-lint.mjs --source-host <host[,host]> [--content-root <dir>] flags a source-host <a href> whose path exists in the content tree as 🟡 D4 LOCALIZE. Advisory by design in this release; promote to 🔴 after one rollout has run the stage cleanly.

0.19.4 — replica: chrome-parity probe (P3)

  • New replica/scripts/chrome-parity.mjs — computed-style parity for chrome. Probes the same regions on live and build (default header + footer; --region strip=<liveSel>|<buildSel> for sticky strips), pairs every text-bearing element by text, and prints only the deltas: font family / size / weight / style / line-height / letter-spacing / transform / colour / background / padding / radius, element rect, the clickable box of links and buttons, plus an icon inventory (count, size, signature) paired by order; MISSING / EXTRA texts on either side. Live-side hardening via the shared live-session.mjs (UA + headers, challenge fail-loud exit 3, overlay dismissal, --headed, --locale). Exit 0 quiet, 2 deltas. Field evidence: one run found what many pixel-band rounds had not (italic-vs-normal note, regular-vs-bold link, wrong nav link colour, 12px row offsets, 97×40 vs 71×32 button, six missing icons).
  • Gate doc, pass bar item 5: styles diagnose, pixels confirm — run the probe BEFORE any pixel iteration on chrome and clear its deltas; iteration discipline gains the same rule. Replica SKILL setup copies the script and the Phase 4 snippet shows the call; deploy Step 10 item 4 points at it for the deployed-origin chrome gate. The ≥98% crop gate remains the pass bar — the probe is a diagnostic, not a new threshold.

0.19.3 — replica field harvest, part 2: row-level instruments, masks, detectors (P1, P11, P22, P16, P19, P15, P20)

Second fold of the 2026-09 same-design-migration ledger: the entries that needed a flag or a small script. Each is additive — a new instrument, a new flag, or a lint/gate DETECTOR; nothing changes what the pipeline emits.

  • Replica — pass bar item 5 (chrome crop gate): chrome means every site-wide repeating band (header, sticky strips, footer); crops are element-anchored PER SIDE (fixed-y crops false-read the moment one side's rhythm shifts — a strip read 66% mis-anchored, 1.6% re-anchored); and authored-volatile regions (campaign heroes) are masked out of the number. pixel-compare.mjs --mask <yA:h[@yB]> neutralises the row band on both sides, removes it from the denominator and prints every mask on the verdict line (P1).
  • Replica — new scripts/row-profile.mjs: (1) a per-column colour-class scan (white/dark/brand/photo run lengths at N x positions) to establish section boundaries from the capture instead of eyeballing crops (P11); (2) brand-colour landmarks — rows dominated by a saturated brand colour, paired live-vs-proto in order, with per-pair delta and gapShift naming the one inter-landmark gap that absorbed a vertical offset (P22). Gate doc § Reading the band breakdown documents both; replica SKILL setup copies it.
  • Extract — crawl.mjs saves the settled rendered DOM as pages/<slug>.html (renderedHtml field): capture once, parse offline; migrate's inputs name it as the structure source for importers (P16).
  • Deploy — Step 3: page templates that cap main > .section > div define ONE full-bleed escape at template specificity; qa-gate.mjs --full-bleed a,b warns when a listed block's section wrapper computes narrower than the viewport (P19).
  • Deploy — D15 lint: inline-script text lifted as copy (window., try {, function () is 🔴; ALL_CAPS_TOKEN tracking lookalikes are 🟡 advisory. Detector only — no capture-time filtering was added (P15).
  • Deploy — block-roundtrip: a dead text whose words are absent from the decorated unit is reported as DROPPED CONTENT (the decoder never consumed that element type) rather than DEAD TEXT; Step 8 names the full default-content set every decorate() must consume. Detector only — no leftovers pass was added to the scaffold (P20).

Deferred by design: P3 (chrome-parity probe), P6 (depends on P3), P12 (sibling variance probe), P24 (link-localization stage + lint tier).

0.19.2 — housekeeping: no site names in the plugin

Every reference to a real customer, test or donor site — in skill text, reference docs, script comments, the deploy IMPROVEMENTS log, the notes folder and this changelog — is replaced by a generic sector descriptor ("a financial-services site, 8 pages", "a fashion retailer's newsletter modal"). Numbers, dates, finding IDs and technical content are unchanged; no script logic was touched (comment lines only, syntax-checked). Kept as is: vendor/platform names (Akamai, Cloudflare, OneTrust, Typekit, Shopify, DA/EDS), Adobe's own properties, fictional sample brands (Wasatch Back, Ledgerline, Evergreen Bank, Meridian Airways), design-vocabulary brands used as aesthetic references, and the eval suite's live crawl target. notes/improvement-plan-2026-08-rwe-centene.md is renamed to notes/improvement-plan-2026-08-replica-deploy.md.

0.19.1 — replica field harvest (2026-09), part 1: one-bullet learnings (P2, P4, P5, P7–P10, P13, P14, P17, P18, P21, P23, P25)

First fold of a 25-entry learnings ledger from a same-design migration of a large vendor-templated site (2026-08/09, published-origin gated, ~150 pages). This part ships only the findings that scored 3/3/3 on general / safe / small — each is a single bullet or an edit to an existing one; no scripts, no new sections. The remaining entries (element-anchored + masked chrome crops, a computed-style chrome-parity probe, sibling variance probing, capture-time code-artifact filtering, rendered-DOM capture, a bundled link-localization stage + lint tier) follow in separate PRs.

  • Replica — recreation procedure: authoring step 1 reconciles component COUNTS from the capture before authoring (vendor templates repeat whole widgets; a big height delta with matching section order is a duplicate, not a missing section) and points at the paragraph-boundary rule; canon chrome is re-verified against EACH new archetype's live page before page content is iterated, with page-level compensation flagged for canon back-port; no foundation text-wrap: balance in replica prototypes/block CSS; the overlay-scrim bullet now reads the full computed background-image layer list FIRST and luminance-fits only when the scrim is genuinely undiscoverable; the sizing-MODEL lift covers heights and overlaps at two or three widths with the vw encoding formula; icons and vectors are harvested from the live DOM (svg.outerHTML), never approximated.
  • Replica — gate: iteration discipline gains the post-pass typography spot-check (a pixel pass at the wrong base metric is latent sibling debt; compensating spacing is the tell); the wide-viewport check samples heights and one intermediate width; the published-origin gate gains two rules — re-probe live chrome metrics at deploy time (crawl captures are the content source, live-now the chrome source; mask authored-volatile regions) and budget one anchors-driven reconcile round at the published origin (the harness number is provisional); the EDS transform list notes that authored inner blocks may flatten, so :has()/block-class selectors are verified against the delivered .plain.html.
  • Migrate — content preservation: paragraph boundaries come from the source's block-level nodes, never from splitting captured text on newlines (inline elements fragment a 5-paragraph disclaimer into 16 <p>s and double the section height).
  • Deploy: favicon delivery is verified with a HEAD at the published origin; Step 3 section styles that paint several wrappers as one surface contain child margins with display: flow-root (margin collapse through an unpadded wrapper paints a ground-colored stripe inside a "card"); Step 6 chrome rows never pair a fixed height with vertical padding under the border-box reset; the Step 7 brief bans <br> inside flex/grid containers (it becomes a sized item) and bare > span-style child rules in variant CSS (they resurrect hidden elements); Step 10 item 5 samples section heights at an intermediate width.

0.19.0 — Experience Workspace editability contract (EW1–EW10) + gate

Every text an author wrote in a DA document must be inline-editable in Experience Workspace (da.live canvas, "quick-edit") once generated block JS has decorated the page — and the block must look the same while it is being edited. Field finding (an energy-company site, 2026-09-03, two rounds: 3 blocks, then 17): over a 29-page covering sample only 841 of 1452 authored texts were editable; every template-slotted block was 0 %. The generated blocks were correct implementations of the skill's own guidance (value-slotting, text(cell), clone-the-anchor) — the guidance was the bug. Mechanism verified against da.live editor-utils.js/prose2aem.js and da-nx quick-edit.js/prose.js; deploy improvement #123. Minor bump: new gate + new qa check.

  • Deploy: § Target runtime documents the workspace instrumentation (data-prose-index on outermost editables, decorate() re-runs over it, only surviving indices become editors). § 2b redefines template-slotted as node-slotting and bans value-slotting. § 3 ships three edit-mode foundation snippets (CTA repaint from <strong>/<em> marks under .prosemirror-editor, card-as-link inner anchor, :where() wrapper variants at equal specificity) + EW10 for section prose. § 5 Buttons, #55, #62/#71, #70 and § Section heads rewritten to MOVE authored elements. § 8 gets a move-based scaffold (wrapNode, labelWrap, stripInstrumentation) and the named Experience Workspace editability contract (EW1–EW10) with the gate command; Step-7 brief carries the contract; Local QA + Checklist gain the EW gate, edit-mode simulation, static review and pixel-parity lines; anti-patterns 18 (value-slotting) and 19 (class on the authored element); References cite the da.live/da-nx sources.
  • Scripts: new deploy/scripts/ew-editability-probe.mjs (URL and --content harness modes, --simulate-editor drift report, @ew-exempt JSDoc tags); block-roundtrip.mjs --ew (default on) fails dead non-exempt texts and duplicated indices 🔴; render-harness.mjs --ew --simulate-editor + hides body > header; section-schema.mjs emits editableTexts per section; content-inventory.mjs exports the outermost-editable classifier and content-diff reports an EDITABLE COUNT advisory.
  • qa: new editability check (editability/dead-text error, editability/duplicated-index warn, per-page summary; --blocks-dir / --ew-exempt for exemptions).
  • replica / rollout / reskin / migrate fidelity-tiers: every block-authoring handoff cites the contract and the EW gate (the brief skipped it on 27/27 blocks because it did not carry it).
  • Evals: new ew-editability (node-slotting, move-not-rebuild, wrapper-descendant selectors, gate evidence, fidelity not traded).
  • Ledger: deploy IMPROVEMENTS.md #123; master reference/learnings.md example entry.

0.18.5 — migration-flow routing: replica subsumes prepare-migration

Routing-surface fix, no pipeline behaviour change. Field finding (an air-cargo site, 2026-09-03): asked "how do I migrate X to EDS with stardust", the agent correctly proposed replica for the keep-the-design route but could not say whether prepare-migration was also needed — the subsumption fact lived only in replica/SKILL.md's Phase 1–5 body, which is never in context until replica is already invoked, and replica was absent from the master skill's routing table altogether. One clarification round-trip per migration conversation.

  • Master skill: routing table gains the missing replica and reskin rows and marks prepare-migration as redesign-flow only. New § Two migration flows — pick one, never mix: redesign (prepare-migration → migrate → deploy/rollout) vs. keep-design (replica → migrate → deploy/rollout, where replica runs extract --prep, a mechanical direction-preservation step, and gated archetype recreation in place of the prep cascade), plus reskin for donor-design/same-content. Instructs stating the chosen flow — and that replica needs no separate prep — in the first response.
  • prepare-migration description: "Redesign-flow only — for same-design migrations stardust:replica runs its own preserve-mode prep cascade; never chain prepare-migration with replica."
  • replica description: "subsumes the stardust:prepare-migration prep cascade in preserve mode — no separate prep step; never chain the two."

Descriptions are the always-loaded routing surface, so the disambiguation now holds even when only the sub-skill frontmatter is in context.

0.18.4 — commerce-site field harvest: chrome crop gate, sizing-model lifts, EDS authoring traps

Harvest of three learnings ledgers from a five-design Magento-PageBuilder → EDS migration (two sibling commerce sites, be/nl + nl, 2026-08, published-origin gated). The headline failure class: small-area, high-salience defects that pass the full-page bar — both pilot runs shipped "green" pages whose header/footer measured only 93–97% match, and a frozen width:720px lifted from an authored width:50% passed both gate breakpoints byte-identically. All changes are site-agnostic; deploy improvements #115–#122.

  • Replica: new scripts/crop-compare.mjs (per-y-band pixelmatch, per-side offsets, default bar 2%); the pass bar gains item 5 — header AND footer bands each ≥98% over the same stitched captures, no extra live hit (#115). New § Wide-viewport fluid check: a ≥1920 box-map spot check catches fluid-vs-fixed width freezes both standard breakpoints render identically (#116). Recreation procedure gains § Lift the sizing MODEL, not the resolved value (two-width lift diff; encode the authored %/vw/max-width rule, never the resolved px; layout models, not wrap outcomes). Iteration discipline gains geometry-fix verification hygiene — rule-bearing element, cache-free serving check (curl --compressed | grep), back-computed reviewer viewport (#117).
  • Deploy: ENCODE contract — never author <hr> (it is the section delimiter; fractures the section at ingestion — lint 🔴, rule HR, #119); rehost assets only from the CAPTURED src and diff dimensions/bytes after fetch (commerce CDNs answer 200 with a generic fallback for guessed paths, #118). Step 3 — one section-metadata style value per section (multi-value delivered only the first class; anchor a second axis with content-scoped :has(), #120); empty-section display overrides must scope to [data-section-status='loaded'] or they defeat pre-load hiding (measured 0.75 CLS, #121). Step 10 gains the chrome crop gate, the ≥1920 box check, and the verification-hygiene items. The deployed computed-style guard also asserts clientWidth > 0 per visible loaded image — loaded ≠ rendered; circular flex sizing collapses an image to 0×0 with naturalWidth still > 0 (#122). sanitise.js now refuses >2 arguments: the two-arg convention made a 3-file batch silently overwrite the second file with the first's content.
  • QA: new zero-size-image check (warn) — loaded image renders 0px wide while participating in layout (getClientRects() guards against display:none false-flags).
  • Reskin: Image-paint gate documents the same loaded-≠-rendered blind spot (paint asserts naturalWidth, not rendered area).

0.18.3 — dual-session field harvest: consent fallback, gate identity assertion, capture-freeze hardening

Harvest of two independent replica+deploy sessions (an energy-company site and a healthcare-insurer site, 2026-08-26/27, on 0.18.2). Three failures recurred in BOTH sessions and lead the release: a stale cross-project :8791 server silently gated a foreign site (once in each direction — every skill doc suggests the same port, so collision on a shared machine is guaranteed); consent widgets missed by the selector list (on the healthcare-insurer site the banner baked into ground truth AND all 7 stitch seams → 32% false pixel diff, one gate round invalidated); and live-data embeds (mirroring the SAME src cancels the data out in the pixel diff — freezing a snapshot guarantees a widget-sized residual). All changes are site-agnostic and additive; the high-impact-but-not-low-risk items (shared-classifier element-boundary separators, stitch-shot --fullpage, per-project default ports) are deliberately deferred with rationale in notes/improvement-plan-2026-08-replica-deploy.md.

  • Extract: crawl.mjs consent dismissal gains a visible-button text-match fallback — exact short labels (Accept / Accept all / Allow all / Agree / OK / Decline / Alle akzeptieren / Accepter), overlay-container scoped, runs ONLY when the selector pass matched nothing, so existing selectors keep priority and an in-content link can never match. Favicon is now captured on the ENTRY page in every mode (bounded --pages extracts skip Phase 3 where favicon capture lived; deploy then skipped silently and shipped the default icon) via an in-page fetch that inherits the context's fingerprint. Bot-wall note: page-level walls usually do NOT gate assets — probe one asset with a browser-UA curl before building in-page-fetch machinery.
  • Replica scripts: gate.sh asserts build-side identity BEFORE any capture — the fetched page must contain a marker (default: the <slug>; --marker overrides), exit 4 names the port listener via lsof; the documented default port is unchanged (the assertion makes collisions loud at near-zero cost). stitch-shot.mjs's freeze now also pauses every <video> at t=0, clears all pending JS timers, and clicks the first slick-convention carousel dot — CSS-only freezing stopped neither video playback (~20% of one page was video noise) nor slick autoplay (slide identity arbitrary per capture; residual 4% → 0.8% once reset). Symmetric on both sides; a static page's capture is byte-identical to 0.18.2's.
  • Replica motion parity (observe, don't infer): new motion-observe.mjs (sibling of stitch-shot, same live-session hardening, exit 3 on challenge) records what the live page actually DOES — animationstart/transitionstart events with element paths + text snippets, class mutations exposing the trigger mechanism, a down+up header-state timeline (dense near the top), --click widget frames, --hover computed-style diffs with the changed-property list precomputed. The interaction-parity pass is now a REQUIRED gate output per archetype — a motion inventory in progress.json (motion: {observed, implemented, dead[]}; live-classed-but-dead behaviors recorded as NOT implemented, the correct replica of a dead class) — because when optional it was skipped on 5 of 7 archetypes and every skipped one shipped visibly static. § Interaction parity is rewritten around the evidence rule: implement ONLY behaviors that measurably fired — static lifting invented motion three field-recorded ways (dead animation classes: 2 of 8 classed caption families ever fired; hover rules whose scope never matches at runtime; approximated chrome mechanisms allowing states impossible on live, e.g. a double-rendered header) — with static CSS remaining the authority for the exact keyframe/easing VALUES of fired animations, mechanisms cloned as the observed state machine, and a two-direction verification (pixel-compare must return to the gated number — field: 1.01% gated → 1.06% with invented motion → 1.01% exact after the evidence-only rewrite — plus a behavior-match assertion off the observe JSON). Full spec: notes/replica-motion-parity.md.
  • Replica docs: two new permanent-residual classes in the capture-state policy (live-data embeds — load the SAME embed same-src on both sides; randomized decorative elements — log, don't chase); AEM-classic richtext byte patterns are load-bearing (mirror them; diff innerHTML when a wrap-count mismatch survives width parity); display: flow-root reproduces clearfix margin containment (fixed −48/−20px per-section errors in one rule); iteration discipline gains the no-op-fix check (an unchanged differing-pixel count means the rule never applied — the round doesn't count); a "verify the port is yours" line wherever :8791 is suggested (also in deploy Step 10 and the diff SKILL).
  • Deploy: Step 3's reset now REQUIRES the global border-box the boilerplate doesn't ship — a bootstrap-era %-width+padding grid silently wrapped every column, +1731px doc height, all text gates green (#106); block DOM must not emit semantic <header> (the stock reservation clamps every one at once, #107); overlay chrome documented as the no-reservation #81 case (--nav-height: 0 + absolute header, measured CLS 0.0004, #108); the block brief requires mobile overrides at the variant's own specificity (#109), flow-root on un-floating overrides (#113), and wrapper resets at lower specificity than the block's own rules (#114); never copy the pipeline's fallback <img src> (750px rendition) into a CSS background — rewrite width=2000 (#110); line-height: 0 on image paragraphs cancels the <picture> wrapper's baseline descender (#111); whitespace-only authored content is dropped by the pipeline — model live spacer line boxes as block CSS (#112); the no-favicon path is a loud WARN recorded in the deploy log, never a silent skip.

0.18.2 — replica field harvest: font-fork instrument fix, interaction parity, published-origin gate

Harvest of a full stardust:replica e2e run (a financial-services site → EDS, 2026-08-25/26, on 0.18.1): a home-page archetype gated to 3.55%/5.56% pixel diff, then 8 pages published and gated against the live origin. All changes are site-agnostic; the validated discipline (measure-first, fail-loud, ≤10% / Δ≤8px / 0-structural-red, hit minimization, no DOM copying) is unchanged.

  • Instrument fix (diff live-session.mjs, F-B2): the standard anti-bot header set now rides DOCUMENT requests only (via context.route), never subresources. Forced on every request it made cross-origin CORS-mode webfont fetches non-simple — they died with net::ERR_FAILED and every live capture silently rendered fallback type, poisoning the whole gate (live doc height moved 6669→6518 after the fix). Bot managers fingerprint the navigation request, which still carries the full set. Companion hardening: stitch-shot asserts fonts loaded after document.fonts.ready and warns loudly on any declared face with FontFace status error (gate doc rule 14 owns the instrument-induced vs capture-state decision).
  • Replica scripts: new anchor.mjs (per-section [y, height] probe — run on both sides, fix the first mismatched section top-down; roughly halved iterations vs band-reading alone in the field) and gate.sh (one pixel round in one command, live capture cached, fail-loud on exit 3).
  • Replica gate doc: new § The published-origin gate — only the published number counts for platform-delivered pages, with the three recurring EDS pipeline deltas (<p><picture> wrapping, empty metadata-section padding, /media_<hash> rewrites); calibration honesty (prototype-regime vs published-origin-regime numbers, same ≤10% bar); probe schedule per fix round (pixels every round, content/visual at milestones — content/visual re-runs cost 2 live hits each); iteration-cap bookkeeping (instrument-invalidated runs excluded once the defect is fixed and named; build-side-only probe passes are free); the script-edit rule narrowed (re-implementing retired adaptations stays a defect; a commented, ledgered, flagged-for-upstream instrument-bug fix is the correct move — fail-loud outranks script immutability).
  • Recreation procedure: CSS lifting gains the text-rendering group (text-rendering, -webkit-font-smoothing, font-synthesis, font-variant-numeric, font-kerning) + the literal-string width diagnostic; new § Interaction parity (hover-diff and behavior-diff probe patterns, Swiper-lock semantics and the scroll-based replica that auto-degrades to the static case); new § Wrap-junction margins (collapsing-margin trap on cards-on-a-canvas sites); capture-state policy gains nondeterministic live elements (tickers, dates, counts — freeze a captured value, log as permanent residual); granularity parity states the widget policy: widgets are implemented, not justified away.
  • Replica SKILL: archetype prototypes are per-archetype and CUMULATIVE (shared canon CSS + per-archetype CSS; never skip to direct platform authoring — prototyped archetypes held 3.5%/5.6% while direct-authored pages plateaued at 8–16%); Phase 5's final proof is now the mandatory published-origin gate.
  • Deploy: preview 409 "error from content-bus" gets a two-step fail-loud diagnosis (known-good doc to the same path, then a per-image sweep for SVGs over the ~40KB hard pipeline limit — rasterize to PNG); #99 extended accordingly.
  • Migrate: sibling content-fidelity is now measured per page at import time — a role-classified node-count acceptance (headings / body / CTAs / images vs the captured page JSON; drops not covered by a logged contentDeviations[] entry fail the page), so dropped-content importer bugs surface while the importer is still cheap to fix.

0.17.0 — vanilla aem-boilerplate is the only deploy runtime; David's Model becomes a mechanical gate

The AuthorKit runtime dependency is removed end to end: stardust:deploy targets stock adobe/aem-boilerplate — no runtime port, no vendored files, no .eslintignore, no pinned author-kit ref. Validated by three full e2e conversions onto a fresh boilerplate + DA site (6/11/13-section pages, root and subfolder scopes); the third run shipped with zero live-only defects.

  • Runtime (deploy): bootstrap-authorkit.mjs deleted; new "Target runtime" contract documents what stock boilerplate provides (never modified). Buttons are the vanilla family (a.button.primary/.secondary/ .accent in p.button-wrapper; the probe records per-target drift — older clones emit button-container). Fonts move to styles/fonts.css + metric-matched <brand>-fallback faces (the stock convention); the body.appear gate is the runtime's and stays. Chrome is authored /nav + /footer documents fed to template-slotted header/footer blocks — nav links become authorable and interactive chrome is real block JS; per-page nav:/footer: metadata replaces header: off (and makes multilingual chrome routing a content concern, no runtime patching).
  • David's Model (deploy): new davids-model.md maps all 15 rules to their enforcement points (D#N citations); the ENCODE contract gains the missing structural rules (D2 nested blocks, D3 spans, D4 URLs, D10 columns, D13 alt-text, D15 code-as-text, D1 auto-blocked embeds); Step 2 opens with D1/D11 triage (prose sections land as default content with a small closed section-style vocabulary; Block Collection patterns mirror collection models). New davids-model-lint.mjs gates the atomic delivery chain — three consecutive e2e runs produced 0 🔴 first-pass content structure with no model instruction in the run prompt.
  • New gates + findings from the validation runs: qa-gate.mjs (stock Local-QA assertion run driven by the page's eds-schema — replaces hand-rolled probes); Local-QA scope boundary (CLS, content-diff/ visual-diff, and chrome overrides are deployed-URL-only checks — the harness false-passes CLS); findings #96–#101 in deploy/IMPROVEMENTS.md, including: the pipeline <p>-wraps nav trigger links (#98), bitmap-embedding SVGs 409 the preview (#99, lint advisory added), and the metadata-first empty section defeats waitForFirstImage so hero blocks must eager-load their LCP image and reserve the media slot (#100).
  • Cross-skill: rollout delivers chrome as published /nav + /footer documents (roster + coverage semantics updated; multilingual routing via per-language documents); diff's BLANK_RENDER hint now points at the runtime not booting instead of advising removal of the stock display gate; edsName() guards -wrapper/-container suffix collisions.

Measured effect across the validation runs: single-page conversion time fell from ~74 to ~42 minutes as findings fed back into the skill, with fidelity gates green throughout (content+roles matched 41/41, 94/94, 71/71 text nodes; live CLS ≤ 0.01 after #100).

0.16.1 — container-width sizing guidance in the token contract

Docs only, no behavioral surface. New ## Sizing --max-width section in skills/stardust/reference/token-contract.md: stardust ships no default container width — inherit the captured container when it holds up (widening one step within the site's own framework vocabulary when it reads dated), measure-first (1200–1280px) when the capture has no measurable container, and persist the derived value to DESIGN.json extensions.breakpoints.containerMaxWidth with the change flagged in the page-shape brief. Cross-referenced from skills/prototype/reference/page-shape-brief.md § Layout strategy.

0.16.0 — two new entry points: replica (same-design migration) and reskin (content × donor design)

Round-1 outcome of the three-new-use-cases exploration (research, candidate designs, and validation evidence in notes/new-use-cases/). Both flows were validated on real pages before codification — replica converged a typographic retail home page to a 1.31% pixel diff with zero structural findings in 3 measured iterations; reskin carried a healthcare site's content byte-identically (2281/2281 chars, 47/47 slots, 13/13 metadata) onto a payments-company donor's token system with 91% of slots mapped to named donor modules. No existing skill was modified (round-2 synergy candidates are listed in notes/new-use-cases/ROUND-1-REPORT.md).

  • stardust:replica (new): same-design migration to AEM EDS. extract --prep unchanged → mechanical preserve-direction (current-state spec promoted verbatim as target; deltas only via the inconsistency register) → clean re-authored archetype recreation (values lifted from the source site's own CSS, never DOM copies) → measured source-fidelity gate per breakpoint (diff's two probes --profile generic + new stitch-shot.mjs/pixel-compare.mjs stitched pixel probe with per-band breakdown, ≤3 iterations) → migrate sibling tier / deploy (template-slotted bias) / rollout unchanged.
  • stardust:reskin (new): byte-faithful content onto a donor design system (live URL via extract --design-source, or local prototypes; Figma donor contract-defined, not implemented). Content-model capture with scope declaration + executable normalization ledger → mapping brief (≥80% slots mapped to named donor modules, no silent improvisation) → programmatic render from the model (never retyped) → dual gates: content (vendored dom-equality.mjs, Apache-2.0 attribution, structure informational + slot-coverage.mjs incl. metadata) and design-adoption (donor-probe.mjs token assertions; selector-missing = FAIL).
  • Both skills were smoke-tested for generalization on fresh sites before shipping (replica: a furniture retailer, desktop converged to 1.06%; reskin: a university site × an analytics-vendor donor, 4883/4883 text bytes, 101/101 slot checks) and hardened from the findings: replica gained pointer-park capture hygiene, the fixed/sticky-chrome × stitched-capture procedure, per-breakpoint CSS lifting, and the full four-patch adaptation set for the diff probes (upstreaming them as diff flags is the recorded round-2 candidate); reskin gained the document-ordered render stream in the content model (ordered + tiling verification), root-kind slot classification, a shared image-visibility predicate across capture and gate, the scope-granularity smell check, and the bounded donor-sampling recipe. Smoke evidence: /Users/paolo/stardust/smoke-{replica,reskin}/SMOKE-REPORT.md.
  • New evals: replica-source-fidelity/, reskin-content-fidelity/.

Field-test hardening (5+5 home pages, findings ledger in the 2026-07 field report)

A 10-site field test (replica: a furniture maker, a luggage brand, a fashion retailer, an EV maker, a fashion house; reskin: botanic-garden×SaaS-donor PASS, museum×messaging-vendor PASS, humanitarian-nonprofit×hosting-vendor) produced an 18-finding ledger; all skill-wrong findings are folded:

  • Shared live-measurement hardening (F-G, F-R1, luggage-brand-1; HIGH). New diff/scripts/live-session.mjs — the one home for hitting live sites to measure them, as robust as extract's capture engine: real-Chrome UA plus the standard request headers (Akamai fingerprints on the absence of Accept/Accept-Language/sec-ch-ua, so UA alone still 403s — reproduced on the humanitarian nonprofit's origin, fixed to HTTP 200; the same header set un-blocked the luggage brand's gate headlessly), challenge detection that fails loud (exit 3, never silently measured as the source), headed-stealth escalation, and two-class overlay dismissal (consent + timed marketing modals, the fashion retailer's #wps_popup case — CH-1). Consumed by diff's two probes, replica's stitch-shot, and reskin's three live-hitting scripts.
  • diff flags replace replica's 10 hand-edits (F-B). --ua, --wait-until, --dismiss, --headed, --locale on both probes and visual-diff --main, backward-compatible for local/deploy use; source-fidelity-gate.md § Script adaptations rewritten — a hand-edited project copy is now a defect.
  • replica: bounded --single entry gets a satisfiable promotion contract (bounded-single synthesis branch — luggage-brand-3); --main body banned with the 103-false-🔴 reproduction (F-C); hit-minimization + media-density iteration budget (luggage-brand-2, CH-2); mobile-@media-first and role-parity recreation guidance (CH-3/FH-2); locale pinning for capture determinism.
  • reskin: ordered stream is now innerText-consistent by construction (F-R2 — the botanic-garden site's a11y ghost labels eliminated at the source; 8/8 orderedVerified vs 5 false in the field) with a sanctioned documented fallback; formControl stream nodes carry select/option/input text verbatim (F-R3 — the humanitarian nonprofit's course form now fully reconstructable, 13/13 verified); slot-coverage gains a paint assertion so an origin-locked CDN can't hide behind a passing URL-string gate (F-R4, the botanic-garden site's 19 unpainted images); zero-output scope errors now guide discovery (F-D); first-match scope semantics and bounded-donor token sourcing documented (F-R5, F-R6).
  • Manifest version aligned (F-A).
  • PR-review P1 fixes (multi-agent review of PR #238): donor-probe expands CSS box shorthands canonically (3-value [t,r,b,r], not cyclic — a pixel-perfect render no longer false-fails the design gate); stitch-shot fails loud on scroll-stall (inner-scroller/scroll-jacked pages can no longer produce silent black-row captures); the diff probes regain their advisory exit contract for HTTP errors (gotoLive httpError:'measure' — a 404 build side reports flags at exit 0 again; challenges still exit 3; reskin's byte gate keeps fail-loud); defaultWaitUntil centralized in live-session with a three-tier rule (localhost and *.aem.page/.aem.live/ .hlx.page/.hlx.live → networkidle, other live → domcontentloaded) so deploy Step 10 never measures a half-decorated EDS page.
  • PR-review P2/P3 fixes: the challenge solve-window runs headed-only — a challenged headless run now costs exactly 1 hit (was 4, the entire recorded Akamai block budget) before exit 3; slot-coverage routes live --rendered targets through live-session like its siblings (challenge → exit 3, no more swallowed navigation errors); case-insensitive stream matching no longer reuses indexes across case-folded strings (Turkish İ class — corrupted stream bytes fixed at the source); bootstrap re-runs preserve the favicon <link> when overwriting head.html (idempotent re-injection); typo'd --flags now error loudly in dom-equality / donor-probe / slot-coverage; the QA harness derives its favicon link from the shipped favicon.<ext> (or keeps the request-free data:, no-op).

0.15.0 — deploy accuracy: close the ENCODE/DECODE round-trip at authoring time (#93–#95)

The six-site e2e campaign showed stardust:diff's structural probe catching real dropped-CTA / role-swap defects on every site — post-deploy, when each fix costs a redeploy loop. Root cause: authored rows (ENCODE) and block decode (DECODE) are written independently and hoped to be inverses. This release moves the defect-finding to conversion time so deploy Step 10 becomes a proof, not a repair loop:

  • #93 section-schema.mjs (deploy, new): the per-section ENCODE/DECODE shared contract — ordered role inventory + repeating-unit groups emitted from the rendered prototype; authored rows and block decode are both written from it (new Step 2b).
  • #94 block-roundtrip.mjs (deploy, new): in-loop per-block gate — decorates the authored content locally with the block's own JS+CSS (no DA, no dev server), diffs the decorated section against the prototype section with content-diff's own classifier, exit 2 on structural 🔴 or on any decorate error (a block that throws or whose inlined JS fails to install must never pass — its raw rows can false-match the prototype). Required per block before deploy, plus one whole-page run before the DA push.
  • #95 decode tiers (deploy): template-slotted (verbatim prototype DOM + role slots — fidelity by construction, for fixed-composition sections nobody structurally edits) vs reconstructive (for authorable repeat groups); tier recorded per block.
  • diff: classifier + differ factored into skills/diff/scripts/content-inventory.mjs, shared by content-diff / section-schema / block-roundtrip so every fidelity gate measures with the same instrument (content-diff CLI behavior unchanged).

0.14.5 — crawler clears Cloudflare managed challenges

extract/scripts/crawl.mjs — the bot-management fallback now validates the probe response, not just that the navigation resolved. A Cloudflare managed challenge returns an HTTP 403 interstitial (cf-mitigated: challenge) without throwing — domcontentloaded fires — so the old fallback (which only fired on a thrown network-fingerprint error) sailed past it and the block surfaced later as a fatal capture-time HTTPError. Observed on a bot-challenged site during the 0.14.4 uplift validation batch, where it required hand-patching the crawler mid-run.

  • Challenge detection at the probe: isChallengeResponse() flags an entry-URL 403/429/503 interstitial (cf-mitigated, cf-ray, server: cloudflare/akamai/edge markers). Either reject mode — a thrown fingerprint block or a challenge response — now triggers the headed fallback; the reason is recorded in _crawl-log.json#discovery.botBlock (fingerprint | challenge).
  • Stealth-hardened headed Chrome: the fallback launches real Chrome with --disable-blink-features=AutomationControlled + ignoreDefaultArgs: ['--enable-automation'] and spoofs navigator.webdriver via an init script on every context (probe + workers — the challenge re-fires per context, no cross-context cookie sharing). fetchTechnique becomes headed-chrome-stealth.
  • Challenge-solve window: clearChallenge() waits for the non-interactive challenge's JS to set its clearance cookie and reloads before validating status — no-op on a normal 200, so zero overhead on the common path. If headed + stealth + the solve window still can't clear it, the run fails with a clear BotChallengeError (interactive solve required) rather than capturing the interstitial as content.
  • Recipe doc (extract/reference/playwright-recipe.md § Bot-management fallback) updated with the two-reject-mode retry rule and the managed- challenge clearing procedure.

Validated end-to-end: patched crawler on the bot-challenged site auto-detects the challenge, switches to headed-chrome-stealth, and captures the homepage at HTTP 200 (2 headings, ~8.9k chars, 9 images); the common headless path (example.com) is unchanged (no fallback, no botBlock).

0.14.4 — Tessl quality pass, part 1 (descriptions)

Description rewrites for the two skills whose tessl-review drag included description criteria: extract (adds a "Use when…" clause + natural trigger terms — analyze/reverse-engineer/capture design tokens — and a "Not for" scraping disambiguation) and prepare-migration (plain-language framing of the prep cascade + trigger phrases + "Not for" migrate/deploy disambiguation). Body text untouched — zero behavioral surface; the conciseness/progressive-disclosure restructuring of extract/deploy is a separate follow-up with its own validation run.

0.14.3 — seventh-site validation harvest (stardust.style) + review fixes

Learnings L1–L9 from the final validation run (full pipeline on stardust.style, hands-off) plus the PR-review findings, folded:

  • crawl.mjs: trailing-slash forms kept verbatim with slash-insensitive dedupe + a guarded 404 slash-retry that records the resolved URL (_crawl-log.json#crawl.slashRetries[]); reducedMotion: 'reduce' on every context + an 800ms post-scroll settle (animated h1s were silently dropped); visible <pre> contents captured as codeBlocks[]; collision-safe slug assignment (query-variant / flattened-path pages no longer clobber one file); sitemap-index recursion (child-sitemap .xml locs no longer queued as pages); page.close() on every exit path via try/finally.
  • Specs: playwright re-probe rule at the start of every rendering skill (--no-save installs are pruned by any later npm i); token-hygiene gate at the FIRST phase commit (master SKILL.md); partial-inventory broken-link carve-out reconciled across content-preservation / migration-procedure / template-and-module-rendering; cinematic sibling handling specced in migrate (assets carried, cinematic-variant-not-consumed recorded); key-facts-in-server-rendered-content ENCODE rule (#86) with the declaration site defined (DESIGN.json.extensions.metadata.keyFacts[]); stale "closed catalog / 5 weaknesses" references reconciled in the master skill, divergence-toolkit, and artifact-map; diff JOIN/SPLIT limitation documented (#87, code fix pending).
  • Versions realigned across plugin.json / tile.json / marketplace.json / README / this file (the #230 drift class).

0.14.1 — six-site E2E hardening (round 1, folded into extract)

Released as part of the six-site validation cycle; the crawl.mjs items listed under 0.14.2's last bullet were folded here first. Documented retroactively — see git history (4a61c83) for the full diff.

0.14.2 — six-site E2E hardening (round 2)

Fixes folded from validating the pipeline end-to-end on six live sites (site D (airline), site E (tools retailer), site A (healthcare), site F (nonprofit shelter), site B (industrial conglomerate), site C (agency)), ranked by cross-site frequency.

  • migrate no longer dead-ends on missing canon (blocking; 4 of 6 sites). The documented prototype → migrate → deploy path never runs prepare-migration, so migrate arrived with no canon and hard-stopped. migrate § Setup now auto-bootstraps canon from the first approved prototype (the prototype --prep write-back, run on demand) when canon is absent and an approved prototype exists; it only stops when there is nothing to derive canon from. (skills/migrate/SKILL.md)
  • bootstrap-authorkit is transactional + refuses the drift-prone default (blocking; 2 sites). Boilerplate removal now runs after the mandatory edits verify, so a drifted/incompatible source leaves the original runtime intact instead of bricking the repo; author-kit@main is refused unless --ref <sha>, --from-sibling, or --allow-unpinned is given. (skills/deploy/scripts/bootstrap-authorkit.mjs)
  • atomic delivery contract now asserts computed layout (silent-failure guard). A .plain.html pass is not a layout pass — the AuthorKit .<name>.block scoping bug ships a stacked single-column page green. The contract's final gate is now a headless computed-style check (grid blocks must compute display:grid, blocks decorated, 0 pageerror) once per template. (skills/deploy/SKILL.md)
  • crawl.mjs, folded in 0.14.1 and confirmed by the runs: five-field _provenance emission, apex→www origin adoption, Usercentrics shadow-DOM consent, <video>/<iframe> capture, and the playwright preflight (--no-save --legacy-peer-deps) + copy-to-project ESM-resolution guidance.

Backlog (single-site or lower-frequency) tracked in the consolidated E2E learnings digest.

0.14.0 — Fable 5 refactor

Design quality

  • Reference-grounded direction. direct researches real-site references via the optional refero MCP (skills/stardust/reference/reference-research.md) before committing to a direction; the curated seed roll is demoted to the fallback when refero is absent.
  • Brand-adjacent refinement tier. A directed middle ground between faithful reproduction and full re-direction, so "polish, don't reinvent" is a first-class target rather than an improvised compromise.
  • Opened catalogs. The uplift/prototype candidate catalogs (what-if amplifications, motion registers) are no longer closed lists: the agent may extend them with evidence-gated entries justified from the captured brand surface.
  • Vision verification gates. extract and prototype verify their own screenshots/renders with vision checks before a step may pass, catching blank captures, broken renders, and layout collapse early.

New capabilities

  • stardust:audit — new skill: a design + SEO + LLM-visibility audit of a site, producing a scored HTML report. Uses the marketing-skills seo-audit / ai-seo methodology when that plugin is installed and built-in heuristics otherwise.
  • Cross-site same-brand extraction. extract --brand-source / --design-source capture brand and design evidence from a sibling property of the same brand, with automatic sibling discovery.
  • Hands-off production mode. skills/stardust/SKILL.md § Hands-off mode runs the full migration chain without conversational gates, folding the previously external master migration prompt into the skills.
  • Run contracts. A per-run learnings ledger plus a stardust/status.jsonl run-status contract, so long runs are observable and each run feeds the next.

Fidelity

  • Runtime-contract detection in deploy/rollout: probe what the target runtime actually serves instead of assuming the authored contract survived.
  • Atomic per-page delivery verify — each page is verified as a unit immediately after delivery, not batched at the end.
  • Foundation-first gate — global foundations (nav, footer, styles, indexes) must verify before page fan-out begins.
  • Link audit across the delivered site.
  • Query-index resilience — index delivery/verification no longer false-fails or silently drops rows on slow propagation.

Performance

  • Parallelism contracts. Concurrent agents coordinate through a state-machine merge-by-slug contract instead of last-writer-wins on state.json.
  • Parallel prototype variants and crawl concurrency in extract.

Fixed

  • Version/reference drift. plugin.json, tile.json, and the README now carry one version, and the impeccable dependency is declared consistently as hard everywhere (tile.json previously listed it as a soft dependency).