Files
faunosteo/.claude/audits/2026-10-30-v1-2-2-reserved-ids-evidence.md

7.3 KiB
Raw Permalink Blame History

V1-2.2 reserved ids: evidence (2026-10-30)

Owner brief « V1-2.2 : correction de sécurité PDF et poursuite A6 » on 2f1f076. Contract .claude/tasks/contracts/2026-10-30-v1-2-2-reserved-ids.md; docs RUST.md §12.9.

1. Audit

  • Written by the engine next to the drawing's ids: osteo-halo only (compose::compose_highlight, PDF view, rasterised as a document of its own, drawing ids kept). Before the fix ensure_halo_filter returned early when the drawing held a filter#osteo-halo; otherwise the filter went into the first <defs> (or a new one placed first), so an element holding the id earlier in document order won url(#osteo-halo).
  • Not colliding: the canvas scopes every drawing id to fv.<view>. (page ids never start with fv.); the annotation layer writes no id, and its data-* cannot be forged (the reader removes every data-*); the filter's result names haloWide / haloTight resolve inside their filter only; class="detail-zone" styles nothing in the app CSS.
  • Historical drawings: 75 committed SVG files; osteo-overlays in 11 of them (legacy editor group), among them supabase/seed/anatomical-sheets/equine/draft/assets/horse.svg (sha256 0ea5202f…, the DEV object). osteo-halo in none but the hostile fixture crates/core/tests/fixtures/svg/hostile/osteo-halo-id.svg. Hence an exact reserved set, not an osteo-* namespace.

2. Pixel proof (Chromium 156.0.8078.4, Image + canvas as browserRasterizer.ts, 1200 × 943, main view, hip selected, halo on)

Composed SVG written by a throwaway integration test (never committed) from the equine drawing with a body injected after the root's opening tag; pixels within ±24 per channel of #b85a3f counted.

Drawing Before (2f1f076) After
equine, unchanged composed, 1602 px composed, 1602 px, SVG byte-identical
+ <filter id="osteo-halo"><feColorMatrix values="0 …0"/></filter> composed, 0 px (marker erased) refused reserved_id
+ <g id="osteo-halo"/><defs/> composed, 825 px (halo lost) refused reserved_id

3. Tests

svg_reserved_ids 9 tests; pdf_session + 1 (the_pdf_refuses_a_reserved_id_and_keeps_its_own_halo); anatomy_compose R5.7 test renamed (the_halo_filter_is_written_once_and_a_composed_root_is_refused, ledger row updated); svg_clean corpus line and svg_sanitiser declared divergence for osteo-halo-id.svg; core-wasm preview.test.ts three spellings on the shipped module.

4. Mutations (2026-10-30-v1-2-2-mutations-reserved-ids.py)

11 of 11 killed: reader check removed, cleaning check removed, exact case only, osteo- prefix, root id unchecked, descendants unchecked, highlight guard removed, canvas guard removed, pre-fix composer restored (guard removed and halo skipped when the drawing holds one), halo not written, reason renamed. A first K9 (« skip the halo when the root has an id ») survived: no tested drawing has a root id, a mutant with no meaning for the fix; replaced by the pre-fix restore.

5. Gates

  • pnpm rust:check: green (fmt, clippy -D warnings on every feature set, 2032 tests summed over the feature passes, ts-rs export, cargo deny).
  • pnpm check: green (typecheck, core-wasm 65, supabase 128, app 566, lint, build). First run failed on r5.census.test.ts: the ledger still named the renamed cargo test; row updated.
  • Module: 559 360 → 560 015 B gzip -9 (+655 B), 1 513 522 → 1 513 968 B raw; Node's gzip 569 350 B (budget 600 000).
  • pnpm check:all on the working tree (code as committed): green, e2e:ui PASSED, app image, worker sandbox tests and runtime image; e2e:shell not run (needs the isolated stack, no change in its scope).

6. Independent reviews (fresh, Opus: the pinned Fable model was out of credits)

  • Verifier: ECARTS(1) LOW, C8: a doc comment in the new pdf_session.rs helper held the shell expansion of id (an unquoted heredoc while writing it), leaking the local uid and groups; fixed before any commit, the diff scanned for other expansions (none). C1-C7, C9, C10 met; floor guard clean on the diff.
  • Security: PASS, no regression or weakened control. Reported, pre-existing, owner decision: MEDIUM root visibility (a drawing <svg visibility="hidden"> with visibility="visible" children reads clean and the appended marker layer inherits hidden: 0 highlight pixels in the PDF, markers gone in the canvas too); LOW root stroke* inheritance onto the point hitbox; LOW the drawing chooses which <defs> receives the halo (sRGB vs linearRGB, cosmetic); LOW intended behaviour change (re-export of an archive or a logo holding the id refused).

7. Root visibility (brief « V1-2.2 : arbitrage définitif A6 et correction visibility », on 81f7639)

  • Red first: error names added, no check wired: svg_root_visibility 2 of 6 failed (reader accepted hidden, composers composed it), pdf_session 1 failed (export accepted it); a third failure was a fixture of mine (a literal tab is not the stored form) and the test now checks the cleaned stored form. Green once the check is wired.
  • Browser proof: RUST.md §12.10 table. Composed SVG by the core at 81f7639 (stash) and with the fix, throwaway integration test never committed; measured with Playwright Chromium 156, Firefox 157, WebKit 27.2 from the local cache (no install, every request outside about:/blob:/data: aborted). The accepted variants' SVG are byte-identical before and after; the three hiding roots are refused after.
  • Mutations: same script, V1 to V9 added; 20 of 20 killed (K rerun, so the reserved-id protections stay effective).
  • Gates on the working tree: pnpm check:all green (cargo 2065 passed / 0 failed summed over the feature passes, ts-rs types unchanged, wasm, core-wasm 65, supabase 128, app 566, lint, build, e2e:ui PASSED, app image, worker sandbox tests and runtime image; e2e:shell not run). First run stopped on an unused import in the new suite (clippy -D warnings), fixed. Fixtures: cairo, Figma, Inkscape, ImageMagick and owner-folder tests green.
  • Module: 560 015 → 560 377 B gzip -9 (+362 B); Node's gzip 569 817 B (budget 600 000).
  • Independent reviews of the working tree (fresh, Opus). Verifier ECARTS(3), all LOW: R3 the stroke properties no marker sets were argued harmless without proof (docs corrected, now an open finding); R7 TODO J / J6 still waited on Illustrator (closed by BDR-090); floor guard: assertions moved into common/sheet_readers.rs (clarification added to the contract). R1, R2, R4-R6, R8, R9 met. Security PASS (semgrep 352 rules, 0 findings), no regression, osteo-halo protections intact; three pre-existing MEDIUM marker-hiding vectors reported for the owner: root viewBox framing a view without view_box (all 19 equine main-view markers out of frame, horse pixel-identical, accepted Clean); inherited stroke-opacity / stroke-dasharray (canvas outlines gone, detail trigger 0 px in the three engines); inherited stroke-width on the point hitbox (13 of 19 equine markers send their click to the last hitbox in Chromium and WebKit). LOW: halo placement in a drawing-chosen <defs> (Firefox, WebKit lose the halo), root systemLanguage / display canvas-PDF divergence (whole view), colour camouflage (inherent). Not built: the class fix changes the R3 highlight oracle and the R4 digests (owner brief: stop before such a change).