adversarial analyze: per-COU recall columns are silently empty on committed corpora (stale spec_path)

[cloudronin]

Summary

uofa adversarial analyze silently emits empty per-COU recall columns on every committed Phase 2 corpus, because batch_manifest.json records spec_path as a generation-time absolute path that no longer resolves.

This is the same defect class as the out_dir staleness fixed in 548224d, in the second of the two paths that manifest records.

Evidence

spec_path values in the committed batch manifests point at the generator's temp directory:

$ python -c "import json,os; d=json.load(open('dev/build/adversarial/phase2/holdout-2026-04-29-v0513/batch_manifest.json')); \
ps=[s['spec_path'] for s in d['perSpecResults']]; print(ps[0]); print(sum(map(os.path.exists,ps)), '/', len(ps))"
/tmp/holdout_specs/ce/w-al-01.yaml
0 / 39

classifier.py (~line 384) uses it to derive the per-COU baseline key:

try:
    spec_path = Path(per_spec["spec_path"])
    spec_obj = load_spec(spec_path)
    baseline_key = _detect_baseline_key(str(spec_obj.base_cou))
except Exception:
    pass

The bare except: pass means the miss degrades baseline_key to None rather than raising, so the run completes normally and the columns come out blank.

Confirmed on a fresh analyze run of the v0.5.13 holdout (post-548224d1, so per-spec resolution now works and rows are produced):

column	distinct values
recall_morrison_cou1	['']
recall_morrison_cou2	['']
recall_min_per_cou	['']
recall_cou_disparity	['']
cou_dependent_flag	['']

All other columns populate correctly.

Impact

• Per-COU recall (the D1 per-COU coverage-delta columns) cannot be derived from any committed corpus.
• The failure is quiet: the run succeeds, the CSV has the right shape, and only the values are missing.
• Anything that quotes a per-COU figure sourced from these corpora needs re-checking against a run where baseline_key actually resolves.

Suggested fix

Same shape as 548224d — derive rather than trust the recorded path:

1. Resolve the spec by spec_id against the repo's own dev/specs/ tree, falling back to the recorded spec_path when it resolves.
2. Alternatively, read base_cou from the per-spec manifest.json if it is recorded there (needs checking; it may not be), which avoids re-loading the spec YAML entirely.
3. Narrow the bare except: pass so a resolution failure is at least visible — either a warning naming the spec, or a single summary line noting that per-COU columns were not derivable for N specs.

Point 3 is worth doing regardless of which resolution strategy wins: a column that is silently empty is harder to notice than one that is loudly absent.

Notes

• Related: 548224d fixed the sibling out_dir staleness by re-anchoring on the batch directory the caller passes via --in. That approach can't go stale on a future rename because it derives the location instead of recording it; the same reasoning applies here.
• The committed corpora are frozen artifacts, so the fix belongs in the analyzer, not in the recorded manifests.
• Not yet verified: whether the M5 corpus (dev/build/adversarial/phase2/2026-04-26, 381 specs) behaves identically. Resolution of spec_path fails there too (paths have the same shape), but a full analyze run over it was not performed.