Record that gh pr checks omits check runs, and correct fully-clean's fifth case

[d-morrison]

Closes #1074

What changed

Two files, both shared/workflow/.

1. fully-clean.md criterion 1 gains a block on gh pr checks incompleteness

The criterion requiring every check to be completed and passing now says that gh pr checks is not a complete enumeration of a head's check runs, and names commits/<sha>/check-runs as the authority for "has everything finished".

It is explicit that this is a different gap from the one the criterion already documents two paragraphs above. That passage warns a workflow run may produce no check run, so a check-runs query cannot see it, and sends the reader to the raw workflow runs. Here the check run exists and the check-runs endpoint returns it; it is gh pr checks that omits it, so the raw-workflow-run cross-check finds nothing missing and confirms the wrong surface.

Per the issue's "Not yet established" section, no mechanism is asserted. The block names the three untested candidates (app filter, required/branch-protection set, in-progress omission) and records the counts that disqualify each without supplying a replacement.

2. The fifth case's "no check run at all / by construction" claim is corrected

That section stated the reviewer contributed "no check run at all" on #1005 and #1008 and generalized to "the check surface is silent about that reviewer by construction". Both observations are now kept with their dates. No claim is made that the behaviour changed -- two dates cannot establish that, and no change is needed to explain the original record, since the two surfaces disagree on those same PRs on the same day.

The section's operative advice (settle it from get_reviews filtered by login) survives and is stated as strengthened, because there are now two distinct failure modes landing on the same remedy from opposite sides: a check run the rollup omits, and a green check run with no review behind it.

3. An incidental correction that runs the other way

The fifth case carried a retraction of a 04:50:41Z figure, calling it an invented particular. That retraction is itself wrong, and it was produced by exactly the gap this PR documents -- a query against the surface that omits the check run, read as establishing the check run does not exist.

metacognitive-monitoring.md asserted the same refuted claim ("#1008 carries no Copilot-attributable check run at all"), so it is corrected in the same pass rather than left contradicting fully-clean.md. Flagging it here because it is outside the issue's stated scope.

Verification

All re-measured on 2026-08-03, not taken from the issue.

PR	head	gh pr checks	check-runs API	Copilot in rollup	Copilot in API
#1056	cbf39b64	10	13	0	1, success
#1005	a17f1021	8	9	0	1, success
#1008	7abfed6b	9	11	0	1, success

• All 13 check runs on Add refresh-claude-token: rotate CLAUDE_CODE_OAUTH_TOKEN without /install-github-app's workflow scaffolding #1056 report app.slug: github-actions, so no app-level filter selects the omitted subset.
• Every run on Add refresh-claude-token: rotate CLAUDE_CODE_OAUTH_TOKEN without /install-github-app's workflow scaffolding #1056 read status: completed at this measurement, so no in-progress filter explains it either.
• The rollup keeps both copies of validate and both of new-line-breaks / check-new-line-breaks, while dropping one of two build / build and one of two claude / claude, so it is not collapsing duplicate names.
• Check run 91327863807 on 7abfed6b is named copilot-pull-request-reviewer, completed_at: 2026-08-01T04:50:41Z, conclusion: success -- reinstating the retracted figure.
• ums: three findings from a five-PR session #1008 has zero Copilot reviews and a green Copilot check run: the green-check-no-review case directly.
• Add refresh-claude-token: rotate CLAUDE_CODE_OAUTH_TOKEN without /install-github-app's workflow scaffolding #1056's two Copilot reviews sit at 252d8fb5 and 1e17d166; its Copilot check went green at cbf39b64, three commits later.

Checks run before pushing

sync-codex-skill-wrappers.py (no wrapper changes), validate-skills.py, check-links.py (947 links, 434 files), check-vendored-drift.py, markdownlint-cli2@0.22.1 (456 files, 0 errors), check-memory-file-size.py, test_check_memory_file_size.py (13 passed).

check-new-line-breaks.py run after committing against origin/main...HEAD: no lines missing semantic breaks. Added lines scanned for banned punctuation: 131 examined, 0 hits.

[d-morrison]

Incidental observation while opening this PR, worth recording because it is the same shape one surface over.

The Copilot review request landed, and the two surfaces disagree about it:

$ gh api repos/Morrison-Lab/ai-config/pulls/1079/requested_reviewers --jq '.users[].login'
Copilot
$ gh pr view 1079 --json reviewRequests
{"reviewRequests":[]}

Not claiming a shared mechanism with the check-run omission this PR documents -- requested_reviewers is a different endpoint, and I have not tested why these two disagree either. Recording it as a second, independent instance of a gh convenience view returning zero Copilot-attributable entries where the REST endpoint returns one. Will fold it into the fragment only if a reviewer thinks it earns a line.

[github-actions]

Claude finished review — View run

---

No issues found. Checked for bugs and CLAUDE.md compliance — this is a documentation-only PR (shared/workflow/fully-clean.md, shared/workflow/metacognitive-monitoring.md) that I fact-checked directly against the live GitHub API (check-run IDs, SHAs, timestamps, counts for PRs #1056, #1005, #1008), and every empirical claim verified exactly. Style rules (no em-dashes, cross-reference integrity) are satisfied.

Verdict

Ready for merge

[github-actions]

💰 Cost: $6.7187 (review) — run

[Copilot]

Pull request overview

Documents gaps in gh pr checks and corrects earlier conclusions about Copilot check runs.

Changes:

• Recommends direct check-run API verification.
• Corrects evidence for PRs #1005, #1008, and #1056.
• Updates the related metacognitive case record.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

File	Description
shared/workflow/fully-clean.md	Revises fully-clean criteria and Copilot review guidance.
shared/workflow/metacognitive-monitoring.md	Corrects the check-run verification history.

Suppressed comments (2)

shared/workflow/fully-clean.md:717

• Optional style fix: this touched paragraph packs several major clauses into each long line, contrary to the one-clause-per-line convention in shared/writing/semantic-line-breaks.md:1-11. Split the whole paragraph semantically in one pass, as that guide requests for touched sections.

The commit check-runs endpoint disagrees with that rollup on both PRs, returning 9 and 11 respectively, the extra entries including one `copilot-pull-request-reviewer` run apiece.
Two figures in the paragraph above have since been corrected, and the correction runs opposite to the one this record previously carried.
An earlier revision claimed Copilot's own check run completed `success` at `04:50:41Z` on #1008, a later revision retracted that as an invented particular, and the retraction was the wrong one: check run `91327863807` on `7abfed6b` is named `copilot-pull-request-reviewer` and reads `completed_at: 2026-08-01T04:50:41Z`, `conclusion: success`.
It is worth leaving the whole chain visible rather than quietly deleting it, because the retraction was produced by exactly the gap criterion 1 now documents --- a query against a surface that omits the check run, read as establishing that the check run does not exist.
The record's other figure moved too: #1008's rollup count was written as 10 and re-measures at 9, and why was not determined.)

shared/workflow/fully-clean.md:102

• This endpoint is paginated and returns only 30 check runs by default. Without --paginate, a head with more than 30 runs can still produce an empty result even when a later page contains an unfinished run, defeating the promised complete enumeration.

   gh api "repos/<owner>/<repo>/commits/<head-sha>/check-runs" \
     --jq '.check_runs[] | select(.status != "completed") | "\(.name) \(.status)"'

[d-morrison]

This PR merged at 03:36:11Z; Copilot's review posted at 03:46:25Z, ten minutes later. All five findings (3 inline, 2 in the Suppressed comments (2) block) are correct, and all five are addressed against main in #1082, commit 372a9f7.

• --paginate missing (suppressed, :102) --- Addressed. Added --paginate plus per_page=100, with a paragraph and a Don't: bullet on why it is load-bearing: an unfinished run on page 2 returns the same empty result as a finished head.
• Endpoint omits legacy commit statuses (:85) --- Addressed. Added the commits/<sha>/status query and reworded the Do: bullet to ask for both rather than either alone. Also recorded that its combined state reads pending with zero statuses.
• Raw-run claim too categorical (:92) --- Addressed. Your particular checked out: run 30781006647 is Running Copilot Code Review at head_sha: cbf39b64, and gh run list --commit does surface it. Reworded to call the raw-run route indirect and best-effort rather than blind.
• "An absent check run" (:695, :713) --- Addressed. Both occurrences now name the rollup as the thing that omits it, and :693 was tightened the same way. This was the sharpest of the five, since the ambiguity undercut the correction this PR exists to make.
• Semantic line breaks (suppressed, :717) --- Addressed. Reflowed one clause per line. check-new-line-breaks passed on it, since it flags multi-sentence lines rather than long single-sentence ones, so CI did not gate this.

Nothing rebutted or deferred.