feat: stable identity map — ownership sidecar, archify locate, drilldown bundles

[tt-a1i]

Problem and behavior

This draft lets an author attach repository-path ownership to diagram components, classify a Git change range against that map, and navigate from an overview into manifest-bound child diagrams. It reports unclaimed and ambiguous paths explicitly. It does not infer runtime impact, recommend review order, or decide whether a PR should merge.

The recovery candidate is cc1b33a7a02351347842dad4ebe9c13278afc166, based on main 6db72a9. The original mixed checkout and index were preserved; all integration was performed in a separate worktree. This PR remains draft and unmerged.

Implementation and compatibility

• Ownership classification keeps explicit exclusions, ambiguous matches and uncovered paths distinct. Empty-segment/brace subset counterexamples, Unicode paths, genuine Git symlink/gitlink entries and divergent endpoint trees have executable regressions.
• Locate validates its receipt and HTML, names missing/invalid map fields with repair suggestions, and preserves earlier outputs on failure. Newly added and renamed maps receive explicit comparison states; comparison failure remains supplemental to the Locate receipt. Import observations remain authored facts, not inferred topology.
• Bundle delivery reuses canonical validation and manifest-bound filenames, including custom sidecars. A moved output directory remains usable after the source is removed. Missing required bundle inputs fail closed; ordinary Locate can omit optional child details while preserving the parent classification.
• Viewer changes are maintained in modular source. Backspace follows the Escape dismissal/return ladder only in active or nested drilldown. Ordinary pages do not register bundle message listeners; a real nested child still handshakes without its own embedded manifest. Offline messaging keeps source-window, role, payload, ID and digest validation.
• Projection state and count badges are removed from export clones without changing live interaction state. The ordinary HTML wrapper changes, so HTML byte identity is not promised; all SVG bytes in the final regenerated examples/cases stayed unchanged.
• README/reference updates document lint ownership semantics, manifest failures and the distinction between requesting descent and a completed child handshake. The named source-string test was replaced with actual receipt assertions. The recovered decision record is historical; six named stale references now use verified fixed-revision symbols.

Broad exclusion presets, a receipt-only output redesign, extra uncovered widgets, an inference engine and automatic PR comments are not included. The reasons and original requirement mapping are in the disposition record.

Validation

• Final handoff contract checks: 44 passed, zero failed/skipped. They include public CLI field failures preserving previous files, import endpoint attribution, optional child ownership and unsupported comparison types.
• Final real-browser message/navigation checks: 6 passed, zero skipped, using Chrome 153.0.8010.36. These cover a genuine cross-origin OOPIF, native Tab leaving the frame, ordinary keyboard behavior, child focus-clear/return, repeated navigation, malformed messages, changed child identity, unsafe URLs with no request, and deletion after parent load.
• Earlier real-browser acceptance also verifies moved-directory child projection after deleting the source and clean SVG export while preserving live state.
• Packaged/development examples, Gallery, Checkout comparison, pinned MCO artifacts and README animation were regenerated. The official Node 22.23.2 ZIP contains 93 entries; only the expected template/examples, Locate CLI and two reference documents changed from the previous archive. generate-viewer --check, actionlint and diff checks pass.
• Full local tests on product commit 0e38aa9: 1,489 total, 1,437 passed, zero failed, 52 skipped; tracked source hashes stayed unchanged. The final delta changes only the Motion Governor browser test and repository acceptance prose. Its focused Chrome 153 / Node 22 run passed 9/9 without skips; all hosted CI gates passed on this head in run 34685055776, including the complete Viewer interaction and bundle/portable-output/export suites. After d83f5ce still failed a real file reload, controlled Linux runs reproduced storage loss from a pre-initialization diagnostic read even in equally large script-free HTML. Without that read, generated Viewers and padded controls retained all twenty reloads per condition. The test removes the interfering diagnostic probe; it retains actual storage-value and Governor-mode assertions on five real file reloads and five same-origin HTTP navigations, plus existing reduced-motion/suspension/storage-error coverage. The acceptance record links both controlled runs and preserves earlier failures. No product storage workaround or specific Chromium internal mechanism is claimed.

Recovered evidence and limits

The experiment directories are offline research assets, not shipped Skill code. Most added lines in the recovery packages are fixed JSON observations and inert historical scoring inputs:

• Fixed boundary replay: 75 observations and six actual CLI runs, including known bad behavior. An independent run produced identical JSON. Six original external map/sidecar files are retained verbatim; the unadopted presets revision is available on an archive branch.
• First-look recovery: 30 original answers, 15 static ground truths and 180 independent metric checks. Main-sample P@3 is 0.767 without Locate and 0.700 with it; all-15 P@3 is 0.711 and 0.644. One model, one trial per condition and non-blind experimenter ground truth do not establish human review benefit. Complete prompts and noise labels are not all recoverable.
• E3 recovery: the 100-step replay matches old JSON byte for byte. Nine commits trigger ten scheduled manual edits; that is not measured maintenance time. The recovered script's 61-anchor selection differs from the report's 67-anchor selection; both are retained, and the recorded report coordinates independently reproduce its complete observations.

No new model or simulated human study was performed. The author's Discussions #368 and #369 were corrected to state these evidence and availability boundaries. Separately reproduced main defects are isolated in #396 (renderer diagnostics), #397 (explicit straight routes) and existing #373 (finite SVG attributes); they are not bundled into this feature. Related nested-reading work #280/#281 is not closed or approved by this PR.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}