test(fixtures): replace out-of-focus annotated_frames with sharper images + re-expand set (post-#170)

[kalwalt]

Summary

Replace the current out-of-focus annotated_frames fixtures with sharper captures, and re-expand the fixture set now that the #170 determinism blocker is resolved.

Background

crates/core/tests/fixtures/annotated_frames/ feeds the absolute corner-error gate (tests/absolute_corner_error.rs, #166 Track A): each fixture is a JPEG of the printed pinball marker plus a .corners.json with the four corners in TL/TR/BR/BL order. The test reprojects each backend's homography corners and asserts max_err <= baseline + 0.5 px.

Two problems today:

1. The images are poor. The fixtures' own README says it plainly: "The current pinball-seq{1..4}.jpg fixtures show what not to do (out of focus → matcher returns matched_id = -1)." An out-of-focus gate image is a weak signal — it exercises the matcher's failure mode more than its accuracy.
2. The set is down to 2 frames (pinball-seq1, pinball-seq4). seq2/seq3 were dropped for run-to-run Rust non-determinism, and pinball-demo for cross-platform C++ divergence.

Why now: the blocker is lifted

The README says the dropped fixtures "will be re-added once #170 lands cross-platform determinism in both backends." That condition now appears met:

• Rust side: HashMap → BTreeMap for Hough vote tally + VisualDatabase keyframes (investigate: Rust backend matched_id is non-deterministic across runs (HashMap random hasher likely affecting BHC tree topology) #170/fix(kpm): use BTreeMap for Hough vote tally and VisualDatabase keyframes to remove Rust-side matcher nondeterminism (refs #170) #171), removing the BHC-topology non-determinism.
• C++ side: unordered_map → std::map via webarkit/WebARKitLib#39, pulled in by the submodule bump in PR chore(submodule): bump WebARKitLib for std::map matcher-determinism fix (refs #170) #172. (Re-confirmed while closing test(kpm): test_full_pipeline_pose fails on Windows with pose component out of tolerance #118, where this same mechanism was the root cause.)

So we should be able to hold more frames stable across platforms than we could in May.

Scope

• Capture new, sharp frames of the printed pinball marker per the README recipe: tap-to-focus, marker filling 40–80% of frame, no motion blur, good contrast.
• Vet candidates quickly with simple_nft_dual (renders both backends' outlines) before annotating.
• Annotate with tools/annotate_corners/index.html (TL → TR → BR → BL).
• Replace pinball-seq1/seq4 and re-add the dropped frames (or fresh equivalents) — target ~4–5 stable fixtures.
• Regenerate the baseline:
  WEBARKIT_REGEN_CORNER_BASELINE=1 cargo test --test absolute_corner_error --features dual-mode -- --nocapture
• Verify the gate passes in normal mode on Linux CI (the gate is Ubuntu-only).
• Consider tightening REGRESSION_EPSILON_PX (currently widened to ~2.0 px to absorb the old drift) — with determinism fixed and sharper images, a tighter epsilon means a stronger gate.
• Fix the README's internal inconsistency (it calls seq1 "rock-solid (sub-pixel)" while the capture section lumps seq1..4 as out-of-focus), and refresh the "why only two frames" section.

Notes

• No code change needed to add fixtures — the test globs *.corners.json.
• Commit .jpg + .corners.json + baseline.json together so the baseline stays in sync.
• Report matched_id + per-backend max-err in the PR body so reviewers can see what's being gated.

Related

• feat(tests): replace symmetric tier-2 gate with absolute corner-error against external references (hand-annotated GT + jsartoolkitNFT-node parity) #166 — the gate's design
• investigate: Rust backend matched_id is non-deterministic across runs (HashMap random hasher likely affecting BHC tree topology) #170 / fix(kpm): use BTreeMap for Hough vote tally and VisualDatabase keyframes to remove Rust-side matcher nondeterminism (refs #170) #171 / chore(submodule): bump WebARKitLib for std::map matcher-determinism fix (refs #170) #172 — the determinism fixes that unblock this
• investigate: DualFreakMatcher tier-2 divergence on pinball-demo at db_id=2 (~14 px > 2.0 px tolerance) #160 — the original divergence investigation
• test(kpm): test_full_pipeline_pose fails on Windows with pose component out of tolerance #118 — where the C++ unordered_map root cause was re-confirmed