Files
faunosteo/.claude/tasks/contracts/2026-10-30-v1-2-2-three-medium.md

4.0 KiB

Contract: V1-2.2 three MEDIUM (marker isolation, click hijack, view frame)

Owner brief « V1-2.2 : arbitrages propriétaire sur les trois MEDIUM », 2026-10-30, on 958228a (branch feature/v1-2-anatomical-sheets). Two lots:

  • Lot A (committed): the frame rule, defect 3.
  • Lot B (working tree, NOT committed, awaiting the owner's authorisation because it changes the frozen R3 references): marker isolation and hitbox capture, defects 1 and 2, with the click proofs.

Diff under review: the working tree (lot A + lot B) against 958228a.

Problems

  1. Both composers (crates/core/src/anatomy/compose.rs) appended the engine's marker layer to the drawing's own root: every presentation attribute the reader keeps on that root reached the markers, inherited (stroke-opacity, stroke-dasharray, stroke-width, stroke-linejoin, fill-rule, color-interpolation-filters on the halo…) or acting on a whole group (opacity, display).
  2. The point hitbox (pointer-events="all", no stroke of its own) inherited a root stroke-width: a giant stroke widened every hitbox and the last one took the clicks meant for other markers.
  3. A view without view_box was framed by the drawing's viewBox: moving it and translating the drawing back left the horse identical and every marker out of frame, accepted.

Criteria

  • C1 Structural isolation, common to the canvas and the PDF view: no attribute of the drawing's root reaches the layer or the halo filter, inherited or group-wide; the drawing keeps its own attributes and renders as before; not a denylist of attributes.
  • C2 Interactive surfaces declare their geometry and capture rule: the point hitbox stroke="none" and a capture on its circle only; ergonomic sizes unchanged; zones and trigger keep their declared stroke.
  • C3 Real click tests in Chromium, Firefox and WebKit (installed engines), touch where testable: normal hock click, giant root stroke, hock after the fix, two close points, invisible and overlapping surfaces; red on 958228a, green after. A missing engine is named, never passed silently.
  • C4 Frame rule in the Rust core, one definition for both composers, the verdict and PreviewSheet: the view's view_box, else the drawing's viewBox read as a browser reads it; preserveAspectRatio handled; annotations entirely in frame, edges included; curves, arcs, invalid coordinates, secondary views, views without view_box, legitimate transforms; explicit refusal with a closed code; the document never rewritten.
  • C5 The reported drawing (horse, offset viewBox, compensating translate) is refused with its 19 main-view markers, the historical annotations unchanged.
  • C6 No frozen reference (R3 oracles, R3 canvas goldens, R4 digests, R4.1 real-raster manifest, step6 goldens, N-1 copies, seed review drawings, equine drawing and fingerprint) rewritten or regenerated; where lot B changes what a reference compares, the change is identified, explained and left for the owner (no updated expectation committed).
  • C7 osteo-halo and root_visibility protections intact; the A6 perimeter unchanged.
  • C8 Gates: cargo on every feature set, clippy -D warnings, mutations, wasm module under 600 000 B gzip, check:all DEV-isolated (lot A), browser proofs.
  • C9 Docs describe only what is implemented (lot A in RUST.md §12.11, V1-2 contract §6.7, ANNOTATION-RENDERING); lot B documented in its patch and the evidence file, not as shipped.

Out of scope

No SQL, no Supabase DEV, no worker DEV, no historical resource change, no merge, no V1-2.3.

Clarifications

  • The junction-point overlap (zones painted and hit after the vertebral junction points, part of their disc records the zone, on the historical sheet in every engine) is pre-existing and needs an owner decision on the capture order: reported, marked fixme in the click spec, not fixed.
  • with_isolated_layer (tests/common/svg_tree.rs, lot B) is the proposed declared change applied to the frozen R3 trees; it is not committed.