evals: fit-cache key settings, coverage.json written before its check, untested evaluation helpers #173

Follow-up from the final Opus review of PR #12 at 4f52036 (findings 2, 3 and 6). None of them changes a committed number: the reviewer regenerated six results byte for byte, and recomputed four triangulation fits that match the cache.

1. The fit cache key leaves out the triangulation settings. fit_inputs in experiments/evals/evals/pose_priors.py (lines 131 to 143 at 4f52036) hashes the poses, intrinsics, pixel threshold and image and depth bytes. It leaves out triangulate.RATIO, MAX_FEATURES, min_points and min_angle_deg (evals/triangulate.py lines 22 to 23 and 170 to 172), so changing any of them silently reuses old fits.

• Fix: add those four values to the spec dict, and add a test that changing each one alone gives a new key.

2. coverage.json is written before its check refuses. main in experiments/evals/evals/coverage.py (lines 868 to 874) writes results/coverage.json before if bad: raise. A failed run therefore overwrites the tracked JSON, leaves coverage.md stale, and feeds map3d.py (line 384), which reads that JSON. The module docstring (line 15) says main refuses to report.

• Fix: check the replica mismatches before writing either file, and write both only on success. Test it with a mocked mismatch.

3. Evaluation helpers with no tests. None of these, and no caller of them, is exercised by a test in tests/ or models/tests/:

• eth3d.visible_scan_points, which decides the ETH3D evaluation points;
• coverage.evaluate_band, the coverage false-observed result;
• coverage.level_world, the vertical cohort's up direction;
• map3d.seen_empty and map3d.space_claims, the Map3D facing verdict;
• drift.reference_scale_check, the GPS rescale.

Each has one right answer on a small synthetic scene, so each needs a strict hand-computed test.