TASK-113 — a check must not read live project state as its expected value

[ranjiao]

What went wrong

Three tests went red on an ordinary afternoon and none was a regression. Each asserted something read out of the project's own state at the moment it was written, so ordinary progress made it false. The same class had been fixed hours earlier in c9018ae (rows_from_store > 20 → > 0, because closing a task removes a projected row).

What changed

1. test_v5_signoff.TestHistoryIsNotRewritten named exactly three V5 closes; three more were signed the day it shipped. It now asserts the property the count stood in for, quantified over whatever the log and journal hold: every signature parses, keeps its disposition headings and provenance labels, and agrees with the counts in its own event. No task id and no total appear in the class.

2. test_one_line_break_rule.TestTheRefusalNamesTheFlag borrowed TASK-038 off the live board. That row closed and left, so the tool answered TASK-038 is not a row on the board and the flag assertion stopped running — a red that misnamed its own cause. The class now builds its own project and its own row under --root, and asserts the row is writable before asserting the refusal, so a vanished fixture and a broken refusal can never be confused again.

3. LOAD-02 reported REL-00 dangling. Its only sources were a signed V5 record recording that the dangling report was empty, plus two spec lines reporting that finding. Writing that a code was gone brought it back: the check was reading its own output as input, and the signed record is append-only, so nobody was permitted to fix the line it named. bin/perry-diagnose now separates a REFERENCE from a REPORT by three structural marks — inside a signed record, inside a blockquote, or in a paragraph naming a check (one of this checker's own finding codes, or a test_… identifier) — and exempted ids are reported beside the count as user_load.dangling_in_reports, never into it. Bare prose saying a code is gone still counts; that boundary is deliberate and carries its own test.

137ffb3's fix (count records, not mentions) does not transfer literally: LOAD-02 already counts one entry per id, so deduplicating buys nothing. What transfers is the move underneath it — ask what the number is for, and count only what the answer admits. bin/perry-diagnose says so where the rule is written.

4. Found by the sweep, same class. test_diagnose's reconciliation test asserted open_decisions == asks.open on this repository. open_decisions sums two registers and asks.open reads one, so the equality held only while no design doc had the open Chosen row the template requires. open_decisions_by_register now carries the queue half and the test reconciles that.

5. A second instance of (3) arrived on the base branch mid-row. A V4 review under perry/evidence/ demonstrates a check going red — "adding one ordinary DESIGN-900-probe.md … turns the test red" — and LOAD-02 charged the project for DESIGN-900. Same defect one document over. The mark now recognises a paragraph naming a test, not only a finding code; and it is scoped to the paragraph because markdown had wrapped that one sentence across four lines.

Broad-sounding, safe for a stated reason: the rule is applied only to ids already undefined everywhere, and only when every mention is a report. TASK-042 is covered by test_foo loses nothing — TASK-042 is defined and was never a candidate. What survives both filters is the demonstration id. A test pins that direction too.

Each check still fires

• V5: test_the_check_fails_on_a_record_whose_dispositions_were_rewritten mutates a copy of the journal in memory and asserts the checker reddens — heading renamed, provenance rewritten, signer blanked, item deleted.
• Flag refusal: disabling the value→flag match in bin/perry-task (the original defect) turns test_two_flags_do_not_produce_the_same_message red.
• LOAD-02: a genuinely dangling id still fires; relaxing split_dangling into always-exempt reddens 4 of the 11 tests in the new class.

Demonstrated, not reasoned about

Merged locally against the four PMO commits the base gained while this row was open — real further closes, a new signature, new evidence — and the full suite is green there: 63 modules, 1864 tests, 0 red.

Not touched

git diff -- perry/ is empty. No signed record is edited.

🤖 Generated with Claude Code