NOLINT inventory audit — ADR-0278 cite-form drift sweep (2026-05-30)¶

Question¶

ADR-0278 (Accepted 2026-05-04) closed the T7-5 sweep with the claim "75 sites, 0 missing ADR or Research-digest references" for NOLINT(readability-function-size). Three weeks later, has that property held for the broader NOLINT surface (every check, not just readability-function-size)?

Trigger: CLAUDE.md §12 r12 requires every NOLINT in tree to carry an inline ADR / research-digest / rebase-invariant citation. Drift accumulates as new fork-touched code lands.

Method¶

1. Worktree chore/nolint-inventory-audit cut from origin/master at bbcaa8d1.
2. grep -rn -E '// NOLINT(NEXTLINE|BEGIN)?' core/src core/test core/tools excluding feature/third_party, feature/iqa/_iqa, and mcp/3rdparty/cJSON (vendored / scoped block-cited at file head).
3. Python classifier: for each NOLINT site, read the preceding 6-line window plus the same line, search for any of ~80 documented justification markers (ADR-NNNN, Research-NNNN, "upstream-verbatim", "C ABI", "atomic_ref", "registry pattern", "extern linkage", "byte- exact", "vendored libsvm", "POSIX feature-test macro", etc.). The marker list is the union of every justification class observed in the already-cited NOLINTs across the tree.
4. For sites that fell out of the classifier, manual context-read of ±10 lines to distinguish (a) regex-miss false positives where the ADR is cited two lines further up than the window, from (b) genuine citation gaps that drifted in after ADR-0278.

Findings¶

• Total NOLINT sites in scope: 222 (NOLINTNEXTLINE / NOLINTBEGIN / inline NOLINT — NOLINTEND markers excluded).
• Cited automatically: 187 sites passed the regex pass on first run.
• Manual triage of the 35 remaining: 18 turned out to be classifier false positives — ADR / "C ABI" / "extern linkage" citations existed but more than 6 lines above the NOLINT (long Doxygen blocks on init / integer_compute_adm).
• Genuine ADR-form gaps: 17 sites required inline ADR cite, of which 1 lives in core/src/output.cpp deleted by in-flight PR #205 (skipped — owner conflict) — leaving 16 sites this PR actually edits.
• In-flight-PR-owned (skipped): 18 NOLINTs total across core/src/model.cpp (16, file deleted by PR #205), core/src/output.cpp (1, file deleted by PR #205), and core/src/feature/sycl/integer_adm_sycl.cpp + integer_vif_sycl.cpp (1 each, both files touched by 19 sibling drafts). Each already carries per-block prose justification; the cite-form sweep on those files will follow once the merge-train upstream settles.

Sites edited¶

Post-edit verification¶

Re-running the classifier with the same widened marker set against the modified tree shows:

• Cited: 204 / 222 (was 187 pre-edit).
• Uncited: 18 sites — all 18 are in files owned by in-flight DRAFT PRs (15 in model.cpp, 1 in output.cpp, 1 in each of the two sycl files). All carry per-block prose justification; the citation form will be applied once the owning PRs land or are closed.

Recommendation¶

• Land this PR (closes the 16 actionable cite-form gaps).
• After PR #205 (drops inert .cpp shadows) lands, run the audit again — the 16 model.cpp sites will be gone; the 1 output.cpp site will be gone.
• After the sycl-touching drafts settle (~19 PRs queued), open a cite-only sweep PR for integer_adm_sycl.cpp:55 and integer_vif_sycl.cpp:57.
• Future enforcement (queued backlog item T7-5b per ADR-0278): a scripts/ci/check-nolint-citation.sh that runs the classifier in CI and gates new NOLINTs lacking an inline ADR token.

References¶

• ADR-0278 — T7-5 NOLINT-sweep closeout (parent cite-form policy).
• ADR-0141 §2 — every NOLINT must cite the ADR / research-digest / rebase invariant.
• CLAUDE.md §12 r12 — touched-file lint-clean rule.
• ADR-0108 — six deliverables (this digest is deliverable #1).
• Audit script (one-shot): the regex + manual-triage Python from this research digest (not committed; rerun with grep -rn + classifier).