test(data): pin that a partially-failed window is never recorded absent

[eaitbrahim]

Follow-up from the independent adversarial review of #271. No production code changes — the shipped behaviour is correct. This closes a hole in the tests.

The gap

repair.py's most dangerous rule is "a gap window may be recorded absent-at-source only if EVERY chunk completed". Getting it wrong permanently writes off a hole the venue was never fully asked about.

test_one_bad_chunk_is_not_recorded_absent_and_a_later_window_still_repairs claims to cover this. It doesn't. It fails the second chunk, so the first chunk lands 299 bars and the surviving window's key shifts from (BASE+1d, BASE+553d) to (BASE+300d, BASE+553d). probed_keys matches by exact key, so the window is skipped regardless of what probed_ok contains — its windows_absent_at_source == 0 and get_gap_probes() == [] assertions are vacuous in their own setup.

The fix

Fail the first chunk instead. Nothing is upserted, the remaining window keeps its original key, and the probed_keys gate is genuinely exercised.

Verified discriminating

With probed_ok.append(window) added to the failure path:

FAILED test_a_window_whose_FIRST_chunk_fails_is_not_recorded_absent

Note the pre-existing single-chunk test_a_failed_fetch_is_NOT_recorded_as_absent does defend this rule for the one-chunk case — so the review's "not pinned by any test" was slightly overstated. The real gap was the multi-chunk path, which is what this covers.

Gates

ruff    All checks passed!
mypy    Success: no issues found in 224 source files
pytest  2757 passed, 1 skipped

🤖 Generated with Claude Code