fix: follow-ups to #290 and #292 (subject-set wording, re-run contradictions, baseline cost)

[MendixMau]

Depends on #315. This branch is built on chore/baseline-word-trim and contains its commits, so merge #315 first. Land this PR together with the exec.sh follow-up PR (fix/exec-followups). That PR changes project-bin/exec.sh's change_lane UI-only text to the same subject-set wording used here.

An independent review of the merged #290 (squash 3a54675) and #292 (squash c1efcb3) found four defects. Refs #274 #288.

1. UI-only lane ignored pages in other modules (#290)

Problem: the UI-only lane in testing-shape.md §3 said "No regression journeys of other modules". A snippet or layout change reaches pages in other modules, and the subject-set table in the same section already said "a page that embeds the snippet".
Fix: both tables now say the journeys of every page the change reaches, including pages that embed a changed snippet or use a changed layout, in any module. The "what does not re-run" column is now "No journeys of pages the change does not reach". The matching text in exec.sh is in the companion PR.

2. "Re-run after any change" contradicted "reviewers never re-run" (#290)

Fix: two places in module-review.md.

• The "nothing grandfathers" line now says the change's subject set is re-run once, by the implementer.
• The Prove (3) "Acceptable" cell now says the implementer re-runs the subject set once and the reviewer verifies that artifact (testing-shape.md §3). "A reviewer re-running what the implementer already ran and recorded" is added to "Not acceptable".

The stage-4 LOOK expiry rule ("re-run stage 4") is left alone: that is the reviewer redoing its own LOOK, not repeating the implementer's tests.

3. testing-shape.md cost (#290)

Fix:

• The Testing by design: UI-only changes get the full gate chain and are re-proven 5-10x per task #274 story is now told once, in a short box in §3. The box keeps its failure examples: other modules' journeys, the reviewer re-running them, and a full mxcli test at every batch close. It also says the Unit default was never set, and that a caption change owes one journey and one look.
• §2's pointer now names that box.
• The full-suite paragraph defers to the event table and keeps "(and the module's whole journey set)".
• Intake Q13 (bin/lib/intake-template.sh) now gives the story as one clause pointing at Testing by design: UI-only changes get the full gate chain and are re-proven 5-10x per task #274. --repair-intake matches Q13 by its heading, which is unchanged. test-bug12-sync.sh passes 61/61, including "the unit-test default question (Testing by design: UI-only changes get the full gate chain and are re-proven 5-10x per task #274) is appended".

Word count: 6,888 → 6,740.

4. agent-roles.md had become baseline at stages 1/2/3/6 (#292)

Fix: the Dispatching section moves to a new skills/dispatching.md (981 words). The new file adds a harness note: forks, resumes, task notices and naming a model per call are Claude Code concepts. On other harnesses the equivalent is a fresh session per task with a written brief.

Routing changes:

• New row dispatching: lead, stages P,1,2,3,4,5,6, baseline.
• agent-roles goes back to its pre-perf(agents): dispatch overrides four harness defaults; shorter session-start ritual (#288) #292 setup row: lead, P,4,5.
• A stub ## Dispatching heading in agent-roles.md keeps old pointers resolving.
• Pointers updated in runbook §1c and ui-loop.md:240.
• Surfaces regenerated by bin/render-routing.sh.

5. PROOF-OF-LOOK text matching #321

module-review.md stage 5 now states the addenda rule. Each re-look adds its own VALID AT: and PROOF-OF-LOOK: lines. Each citation is checked against the nearest stamp above it. A report is stale only when the model has been built past its newest stamp (by commit time).

§ "The final LOOK" now requires:

• a VALID AT: stamp (an unstamped final LOOK is FAULT);
• one PROOF-OF-LOOK: line per page;
• it is not owed until gate-check's derived ## Current stage readout reaches 6, and a hand-written stage line keeps it owed.

This text matches the checker in PR #321 (obligation-check.sh).

Numbers (bin/render-routing.sh --check)

	before (#315 head)	after
Baseline tier	76,259	76,326
lead, stage 1	29,275	27,216
lead, stages 2, 3, 6	31,079	29,020
lead, stages P, 4, 5	27,367 / 31,079 / 29,171	27,452 / 31,164 / 29,256

The baseline rose by 67 words. Most of that is the #321 PROOF-OF-LOOK text added to module-review.md (+89) and the restored failure examples. Before those two additions, this PR alone came to 76,191.

Checks run

• tests/wave2/test-render-routing.sh bin/render-routing.sh: 5/5 on most runs, with 1 intermittent failure (see the flake below).
• tests/wave2/test-baseline-pack.sh bin/render-routing.sh: 16 ok.
• tests/wave2/test-exec-change-class.sh project-bin/exec.sh: 20/20.
• tests/wave2/test-ritual-slim.sh bin/sync-project.sh: 17/17.
• tests/wave2/test-bug12-sync.sh bin/sync-project.sh: 61/61.
• Passed: check-scripts, check-portability, check-no-private-citations, check-no-client-data, check-pr-discipline origin/master, render-routing --check (in sync, under budget).

A flake that was already there, not caused by this PR: under heavy machine load, render-routing.sh --check sometimes reports a random routed skill as UNROUTED, and a rerun passes. It is being fixed separately in fix/render-routing-pipefail-flake.

Field run

None owed for an instrument: this is skill text, routing, and one intake text body. No instrument logic changed.

🤖 Generated with Claude Code

https://claude.ai/code/session_01VJgWP5vEoAsNsJYqCDMGNw

[MendixMau]

Landed on master as part of the integration squash #329 (commit 31e90ae), after independent review — its Review: line is in that commit message. Closing as superseded.

---

Generated by Claude Code