diff options
| author | Alexander_Kabui <alexanderkabua@gmail.com> | 2026-10-07 21:09:59 +0300 |
|---|---|---|
| committer | Alexander_Kabui <alexanderkabua@gmail.com> | 2026-10-07 21:09:59 +0300 |
| commit | e8279e59992c74e2d1d131b56c7bbb4741934494 (patch) | |
| tree | c73dde5226ee06fdde29d6fddb0235edbc616a7a /test/python/test_mapping_methods.py | |
| parent | 85b63abcfdbae235d292a6de4bd47ffe6ba6dafd (diff) | |
| download | genenetwork4-e8279e59992c74e2d1d131b56c7bbb4741934494.tar.gz | |
* tests/search_parity.rs (search_parity_replay): grade clause (a)
full-set equality at the ENGINE seam; keep status classes, error bodies and clause (b) top-10 at the ROUTE level (harness v2). The route gained hits.truncate(limit) in e912c87 (documented limit contract: default 20, cap 100), which blocked the floor at 85b63ab: 36 shapes DIFFed with rec=200/<full> gn4=200/20 missing=<full-20>. Root cause was never the route: route-level full-set grading only ever worked because the calibration stub serves the full recorded body regardless of the forwarded per_page โ real gn3 caps server-side. The truncate is the route-side piece of the per_page invariant (PORT-DESIGN ยง5: the engine retrieves the full match set; the route keeps its limit logic), so set-equality moved below it, to the seam. Match200 grading now: replay through the route exactly as before (200 check, top-10 clause (b) off the truncated response โ first-ten-in- recorded-order survives the truncate; upstream-call accounting), then call genenetwork4::search::search(q, type, ENGINE_FETCH_LIMIT = 1_000_000) directly, in process, against the same stub gn3, and grade the returned Vec<SearchHit> (dataset, name) identifier set against the fixture for clause (a). A match shape now makes exactly 2 stub requests (route + engine); error/surface shapes are unchanged. replayed_hits reports the engine-level full-set size, so the report line keeps its meaning (Q01: rec=200/1197 gn4=200/1197 missing=0). Floor after the fix: 54/54 PASS โ match200 49/49, clause-(b) 36/36, min Jaccard 1.000, report written (target/parity-reports/). Negative controls (fixture mismatch RED/restore GREEN, truncated stub body RED/restore GREEN, top-10-breaking stub order RED/restore GREEN) ran on this commit via a throwaway uncommitted stub hook; evidence in ~/spikes/harness-fix/HARNESS-FIX-REPORT.md and logs/.
Diffstat (limited to 'test/python/test_mapping_methods.py')
0 files changed, 0 insertions, 0 deletions
