fix(ci): bound the identity-ladder harness so the fast lane cannot time out

[abrichr]

Problem

tests/test_identity_ladder.py::test_harness_zero_false_accept_all_configs intermittently exceeds its 900s pytest-timeout in the required fast test job on main (runs 30750630903 and 30760532379 failed on 2026-08-02; 30757998188 passed between them). Pre-existing intermittency, unrelated to the day's merges.

Root cause

Two compounding, runner-speed-scaled costs:

1. Redundant recorded-frame OCR. _verdict rebuilt the recorded anchor — including a full-frame RapidOCR pass over the recorded frame — for every one of the ~22 verdicts per pair, though the anchor depends only on the recorded frame and the substrate flags. ~300 of ~320 full-frame OCR passes recomputed identical results. Measured on an M2 Max: ~0.7s per full-frame OCR, ~0.65s per Playwright render → ~350s total; a 2–3× slower shared runner crosses 900s.
2. Unbounded corpus in the fast lane. The sweep grows linearly with COLLAPSE_PAIRS (14 pairs × 22 verdicts × 5 configs), so every corpus addition pushed the fast required gate closer to its budget.

Fix (scheduling/budget only — the invariant is untouched)

The never-false-accept invariant, the production tier stack (Replayer._verify_identity), the five substrate configs, and the verdict logic are all unchanged. Only harness scheduling changes:

• Anchor caching: each recorded anchor is built once per (pair, substrate) and the recorded frame's OCR lines are shared across substrate variants. Identifier crops get per-pair filenames (idcrop_<label>.png) so cached anchors never read another pair's crop from the shared bundle dir.
• Bounded fast-lane run: the fast suite now runs test_harness_zero_false_accept_bounded_configs — the first pair of every (glyph_class, flank) collapse class (6 of 14 pairs: O/0 and l/1 × digit-flanked/alpha-flanked/purely-numeric) through the same production stack under all five configs. The test asserts the subset covers every class in the corpus, so growing COLLAPSE_PAIRS cannot grow the fast lane's work unless a genuinely new collapse class appears.
• Exhaustive sweep moves to the slow lane: test_harness_zero_false_accept_all_configs now runs where the slow lane already runs — the nightly/dispatch full matrix — gated by OPENADAPT_IDENTITY_LADDER_EXHAUSTIVE=1 on the canonical Ubuntu leg. A new CI-contract test (test_exhaustive_identity_ladder_corpus_runs_in_the_slow_lane_only) pins the flag to exactly that leg: the exhaustive sweep can neither silently stop running anywhere nor creep back into the fast lane. The existing macOS deselect and its contract test are unchanged.

Measured

run	before	after
fast lane (M2 Max)	~350s, unbounded	98s, bounded by construction
exhaustive (M2 Max)	~350s	230s

Both runs assert the identical zero-false-accept invariant; both pass. ruff check openadapt_flow, ruff format --check openadapt_flow tests, lenient whole-package mypy, tests/test_identity_ladder.py (22 passed), and tests/test_ci_workflow_contract.py (7 passed) are green locally. The bounded harness exercises the same safety-path code paths as the exhaustive one, so the 85% safety-path coverage floor is unaffected.

🤖 Generated with Claude Code