feat(tri pr ready): --why compares what a pre-existing failure printed

[gHashTag]

What

tri pr ready --why: a failure called pre-existing because a check of the same name is red on master (or a merged PR, or its workflow's newest master run) is now also compared by what its failing step printed. A different step, or a line here whose shape is not printed there, is NEW REASON -- its own verdict, exit 7.

Closes #5852
Refs #5837 #5839 (stacked on #5839: base is feat/tri-pr-ready-workflow-baseline)

Why

The Corpus ratchet is red on master for four conflicted type names. A pull request that adds a fifth printed the same four ##[error] lines and read as "pre-existing". The difference is in the step's output above the annotations.

How

• Jobs API -> the first failing step of this PR's job and of the job the baseline used (the master walk's check-run, the workflow fallback's run, or the first merged PR failing that name). Logs API -> that step's output, ##[endgroup] to the first ##[error].
• Masked: timestamps, colour codes, durations, shas (7-40 hex with a digit and a letter), ids of 9+ digits.
• Each side's last 60 lines are looked up in the other side's whole output (a dropped line earlier does not shift the window). A line not found as itself is looked up by shape (digits read as #) and printed as ~ here / there, not judged: a count moves when the cause does not.
• Precedence: WAIT 2 > CANNOT TELL 3 > DO NOT 1 > NEW REASON 7 > safe 0. --merge refuses on 7. A comparison that cannot run (not Actions, log expired) is listed as not established and leaves the verdict alone.

Measured (real runs, 2026-10-04, against the committed binary or the same code)

tri pr ready 5781 --repo gHashTag/t27 --why:

  Corpus ratchet (expected-failure ledger)
      also failing in 5 other place(s) — pre-existing
      why, here:  step `A type name may not gain a second definition`, job 111256330898, 5 output line(s)
      why, there: step `A type name may not gain a second definition`, master at 6e3322918, job 111324093808, 8 output line(s)
      NEW REASON -- the same name, failing differently:
        +     + ModuleInterface  NEW conflict
        4 line(s) there are not printed here
        ~ here:  ledger 77 name(s), observed 78
          there: ledger 77 name(s), observed 81
...
VERDICT: NEW REASON -- 1 failure(s) are red elsewhere too, but not for the same reason:
  - Corpus ratchet (expected-failure ledger)

exit 7.

#5663 (one of master's four names, observed 78): SAME REASON, fewer -- its last 5 line(s) are printed there too, the 78/81 line printed under (1 of them with other numbers -- read, not judged:), exit 0. The first draft compared counts as words and would have called this NEW -- found by scanning the 24 open PRs with a failing ratchet before trusting the rule.

#5812: NEW REASON, the failing step differs: Run the corpus ratchet here, exit 7. #5849, #5839, #5828, #5824: every comparison SAME (master walk, merged-PR and workflow-fallback paths each exercised on #5839). About 85 s per PR.

Tests

• 6 new in prcheck::why_tests, on a fixture shaped like job 111329604306's log; the verdict test now calls the real verdict_code. cargo test -p tri prcheck: 59 passed.
• 16 mutations, each turns the suite red: no timestamp strip, no colour strip, no duration / sha / id mask, sha mask without the digit rule, script echo kept, annotations kept, tail compared to tail, step ignored, fewer read as new, 7 above 1, 7 never set, no dedup, numbers judged as words, renumbered counted as new.
• tri census pin --gate: fetches 72 -> 74 lines, 31 -> 33 fetch sites, --paginate 9 -> 11 (failing_jobs_on, failing_job_in_run), blessed in this commit. tri gates tests --gate and tri now check pass.

Not established

• That the same text is the same cause; a NEW REASON is two outputs to read, not proof this change caused it.
• A step whose whole reason is a count: two runs differing only in a number read SAME (the numbers are printed for a person).
• Output after the first ##[error]; a reason only in an artifact or summary; logs past retention (read as cannot compare).
• The full cargo test -p tri was not run locally (its fpga Lean tests clone mathlib4); CI runs it.

🤖 Generated with Claude Code

[github-actions]

📓 NotebookLM Notebook linked to this PR

• Notebook ID: b83263109fb055dc
• URL: https://notebooklm.google.com/notebook/b83263109fb055dc

This notebook contains session context, decisions, and artifacts for this work.

[github-actions]

📓 NotebookLM Notebook linked to this PR

• Notebook ID: b83263109fb055dc
• URL: https://notebooklm.google.com/notebook/b83263109fb055dc

This notebook contains session context, decisions, and artifacts for this work.