Adopt the corpus that audits its own canon and delivers each section whole

[m2ux]

Summary

Definitions and code sit on separate branches, so a merged corpus reaches the code branch only when a commit here says so. This is that commit for 875a36dc, which is three merged pull requests on workflows:

• #582 — a canon audit that took restated Protocol phases and schema-owned routing out of rule bodies, and stopped a Capability being read through a link.
• #586 — checkpoints that asked for an acknowledgement and did nothing with the answer, plus a scope round that now advances before the gate carrying it.
• #588 — section delivery: ten cross-section references become anchored links, one operative framing paragraph becomes a section, and three citations move to the section their step reads.

Three files: the submodule pointer, the corpus stamp, and the delivery-cost fixture.

The walk baseline does not move, and that is the result

tests/e2e/option-coverage.json is unchanged, which is worth stating rather than passing over. The corpus gained checkpoint and gate changes — a task cycle that now stops when a symbol's provenance is unaccounted for, an isolation gate whose abort actually aborts, a gate that no longer asks a question with one answer. Reachability was the thing those could move, so the full fourteen-workflow walk ran: nineteen minutes, and every declared option is still reached by some walk.

The walk reported one failure, and it is an artefact of the run rather than of the corpus: the stamp-freshness assertion read corpus-sha.json at the moment the walk started, a minute before npm run baseline:stamp rewrote it. That assertion compares two strings and walks nothing; the coverage assertion beside it passed on its own evidence.

Delivery cost falls

	Reference	This corpus
Delivery chars	1,394,470	1,392,659	−1,811 (−0.13%)
get_activity	562,073	560,462	−1,611
get_workflow	112,548	112,376	−172
get_technique	147,495	147,450	−45
get_resource	572,354	572,371	+17

The first recording to fall. Almost all of it is prose the canon audit removed, delivered again in each activity that binds the operation holding it.

The single increase has a single source. work-package/resources/findings-report.md is the only resource this walk reads that changed, and its finding-layout section now names the field order as a link instead of saying "above" — seventeen characters that a reader given only that section could not previously follow. The section-grain citation changes in the same corpus pull request are in workflow-design and plain-language, which this walk does not visit, so none of them appears here either way.

Gate re-run against the re-recorded fixture: PASS at 0%.

The triage stamp, and what re-affirming it cost

The binding-fidelity verdicts were made against 131e2942, seventeen corpus commits back. A clean guard run on that much drift says the verdicts are old rather than that they are right, so the question is which of the sixty-four entries could have gone stale — and that is a diff, not a re-read of all of them.

Sixty-five corpus files changed. Exactly one of them is a triaged site: meta/techniques/workflow-engine/finalize-activity.md. Its change is the removal of a parenthetical citing a Protocol phase by ordinal, which touches neither the declared output nor who consumes it, so shared-op-return-contract still reads correctly. The stamp moves to 875a36dc with it.

Verification

• Guard suite: 34 pass, 0 fail, 0 unmeasured.

• Binding fidelity: 64 triaged as accepted debt, 0 live, 0 untriaged, 0 stale.

• Option-coverage walk: every declared option reached, no exemption-list change.

• npm run typecheck: clean.

• Delivery gate: PASS at 0% against the fixture this commit records.

• Unit suite (npm run test:ci): 1,139 passed, 2 skipped, 0 failed — after one expectation moved with the corpus. The comment below has the detail.

Scope

The pointer, the stamp and the fixture. No source, schema or test-logic change.

[m2ux]

The Verify failure, and the fix

One test failed: tests/mcp-server.test.ts — states each option's consequence from the workflow graph before the user chooses. It expected the provide-input option at submit-for-review's body-non-conformant gate to carry { exit, next_activity }, and got those plus ends_activity: true.

That is the adopted corpus, and it is intended. The decisionless-checkpoint work added immediate: true to submit-for-review's provide-input exit, on the reasoning its commit gives: end the sequence where a gate's answer says to leave. Answering that gate means the run leaves and re-enters with a corrected body, so the eleven steps after the gate should not run — and the consequence the checkpoint states before the user chooses should say so.

The abort exit at the same gate was already immediate and its assertion was already right, which is why only one of the two moved.

Fixed by reading the expectation to match: provide-input now asserts ends_activity: true, and the comment above the shared step manifest says both of that gate's exits leave rather than implying only abort does.

npm run test:ci: 1,139 passed, 2 skipped, 0 failed.

Why this reached CI rather than being caught locally

The pull request body says the unit suite was still running when this opened. It was not going to finish. npm test is vitest in watch mode; npm run test:ci is vitest run. I started the first, watched it sit idle for hours, and read that as a long suite rather than as a suite that had finished and was waiting for a file to change.

So the claim in the body — that the result would follow in a comment — was reporting on a signal that could not arrive. The suite is run above, and it passes.