fix(diagnose): LOAD-02 counted the record written to close its own row

[ranjiao]

The defect

bin/perry-diagnose's dangling-id check (LOAD-02) reported ['DESIGN-900', 'REL-00'] on this repository, and every live mention of both ids was inside perry/evidence/2026-08/TASK-113-dispatch-2026-08-20-1813.md — the record written to close the row that fixed this exact check (lines 11, 16, 18, 35). Every other mention in the repository already classified correctly as a report.

So the row could not be closed without writing a record, and writing the record reopened it. The defect is in the rule, not in the two ids, and no wording of a true account gets them out.

The fix — spec option 1, a fourth structural mark

Argued in the same block as the existing three, in the same voice:

in a document that reports on a check, a mention of an id the project has already reported on.

The first three marks ask whether this line is a report; the fourth asks whether this id has one. That cannot be decided from one file, so it lives in split_dangling (now two passes) rather than report_lines. The document half reuses names_a_check read over the whole file — no second vocabulary.

Both halves are load-bearing and each keeps the other honest:

• drop the document half → any id ever named beside a check goes silent everywhere, so burying one live reference under one report silences a true finding (test_one_live_reference_is_enough_to_keep_an_id_dangling);
• drop the id half → every document discussing a check exempts every id in it, which for this repository is perry/evidence/** under another name.

Option 2 (a distinct fourth outcome) was rejected: it needs the identical classification work and would additionally require rewriting the assertion in test_perrys_own_repository_reports_the_exemption_it_used. A wholesale perry/evidence/** path exemption was considered and rejected — it fails the spec's second anti-vacuity direction outright.

Verification (V3)

1. perry-diagnose --only=user_load --root . → dangling: [], dangling_in_reports: ['DESIGN-900', 'REL-00', 'ZZZ-404']. Before: dangling: ['DESIGN-900', 'REL-00'], dangling_in_reports: ['ZZZ-404'] — exactly two ids moved, nothing else changed.
2. Anti-vacuity, both directions, re-run rather than asserted:
   • a temp project whose only content is Blocked on ZZZ-404 until Friday. → dangling: ['ZZZ-404']
   • an evidence/-shaped dispatch record that does name LOAD-02 and a test_ name, and is genuinely blocked on two ids → dangling: ['QQQ-77', 'ZZZ-404']. A wholesale evidence exemption would have printed [] here.
3. The mark is load-bearing: neutralising document_reports_on_a_check reddens both named tests plus two of the three new ones (4 failures). Restored → green.
4. Both named tests pass. No existing assertion was changed or weakened; three tests added.
5. python3 tests/parallel -j 4 → 65 modules · 1933 tests · all green (own baseline at 13cfe2f: 65 modules · 1930 tests · test_diagnose red, 2 failures). bash tests/run green. python3 bin/perry-lint exit 0 with the same 3 pre-existing NS-01 warnings as baseline. git diff --check clean. git diff -- perry/ empty.

🤖 Generated with Claude Code