Make --verify check emptiness and walk disk-to-manifest (#244)

[realmarcin]

Closes #244.

#244 named three gaps beyond the manifest→disk existence check.

One is moot

It asked that the citation sidecar be asserted alongside each artifact. #388 deleted all 353 as a broken regex over report prose — exactly what #249 predicted when it said to decide that issue first.

Existence is not non-emptiness

.exists() passes for a zero-byte file, so a truncated write, a process killed mid-flush, or a provider returning an empty body all read as success.

The floor is set from the corpus, not guessed: the smallest real report is 20,785 bytes, so 1 KiB leaves a 20× margin and cannot fail on real data.

Disk → manifest

A report with no ok row was invisible, while still suppressing the call that would have produced it — resume keys on the artifact existing.

Blocking from the start, with one named exception (cellulolysis-deep-research-codex.md, #245) using the same escape idiom as audit_biolink_curies.py's ALLOWED_UNBACKED, so adding one is a reviewed change rather than a cell edit. It's harmless today only because its -codex suffix isn't the -falcon name resume looks for — luck, not design.

Scoped to .md deliberately: a -meta.yaml from --dry-run also lives under research/traits and represents no research at all (status: dry-run, cost: None, task_id: None — #246). Counting it would let a plan nobody paid for satisfy an existence check.

Found while probing, and worth more than the feature

The counters were per row. The manifest is append-only, and 342 of the 353 artifacts carry two ok rows apiece — the original sweep and the re-run after its output was lost. 700 rows describe 353 files, so a single truncated report was reported as two.

Both counters now dedupe per artifact, which is how the invariants are phrased and how the adjacent CURIE scan already reported.

Testing

ok_outputs / missing_artifacts / undersized_artifacts / orphan_reports are extracted as pure functions in scan_malformed_curies' style, so the gates can be tested at all. Eight tests, each making a gate fire — on the argument that file already states:

A gate whose corpus is already clean is exactly the kind that can be silently broken — a mistyped pattern would keep reporting zero forever. These tests are what distinguishes "found nothing" from "cannot find anything".

Probed by hand before the tests existed: an injected orphan and a truncated artifact were both caught, tree restored clean.

manifest ok rows with a missing artifact: 0
ok artifacts below 1024 bytes: 0
reports on disk with no ok manifest row: 0 (1 known, excluded)
reports carrying a malformed CURIE: 0 (0 matches; scanned 354 artifacts)

just qc green · 548 tests pass (8 new) · ruff clean · history record per #325.

[culturebot-reviewer]

The two gates are right in substance, and the per-artifact dedupe is a genuinely good catch — I confirmed it against the tree: 700 ok rows over 353 distinct outputs, 354 .md on disk, smallest real report 20,785 bytes (ecology/biosafety_level_4), and the one orphan is exactly cellulolysis-deep-research-codex.md. The floor and the exception list both check out. One real problem with the orphan gate's scope.

🟡 The orphan gate blocks far wider than its own rationale, and will turn just qc red on documented workflows

orphan_reports globs every *.md under research/traits/, but only run_trait_graph_audit.py ever writes reports/trait_graph_audit_manifest.tsv (I grepped — nothing else touches it). Two documented single-trait paths write reports into that directory and record nothing:

• just research-trait <cat> <slug> --provider openai — README:122 documents this exact override; research_trait.py:271 writes {slug}-deep-research-openai.md.
• just research-trait-edison <target> — justfile:429, writes {slug}-edison-{job}.md (research_trait_edison.py:242,261) on a real run.

The first non-falcon research run after this merges leaves the tree failing qc, and the only remedy the design offers is editing KNOWN_ORPHAN_ARTIFACTS — a code change per artifact. That is precisely how the -codex file got there.

What makes this a scoping bug rather than a policy call is your own argument. The stated invariant is that an unrecorded report suppresses a call that was never paid for, because resume keys on the artifact existing. But resume keys on output_path() → {slug}-deep-research-falcon.md. You say this yourself about the exception: it is harmless because -codex is not the name resume looks for. That reasoning generalises — -deep-research-openai.md and -edison-literature.md cannot suppress anything either. The gate currently blocks on files that are structurally incapable of causing the harm it exists to prevent.

Narrow the glob to the resume namespace (*-deep-research-falcon.md, or whatever output_path resolves to) and the gate matches its invariant exactly — and cellulolysis-deep-research-codex.md stops needing an exception at all, which is the better outcome for #245 too. If you'd rather keep the wide net, then the non-resume namespaces need to be reported without failing, not exempted one filename at a time.

🔵 Minor

• justfile:908 — the recipe comment still reads "every manifest ok row's report is on disk, and no artifact carries a malformed CURIE". It now also checks emptiness and walks disk→manifest. This is the stale-prose-beside-a-changed-gate failure docs/WORKFLOW_CONVENTIONS.md names, and audit-justfile-paths can't see it.
• manifest ok rows with a missing artifact: (line ~261) now counts artifacts, not rows — that dedupe is the PR's own headline finding. The adjacent CURIE line was reworded for exactly this reason ("reports carrying a malformed CURIE: N (M matches; …)"); this one wasn't.
• undersized[:20] and orphans[:20] print without the ... and N more tail that missing and bad_curies both carry. The totals are on the count line so nothing is truly hidden, but the three blocks now read inconsistently.

Tests, history record (history/infrastructure/verify-artifact-integrity/…) and CI are all in order; the eight new tests each make their gate fire, which is the right bar for a gate that reports zero.