feat(rules): make verification confirm the claim, not the existence of a token

[paveg]

What

Two always-on rules learn to verify the claim rather than the presence of a token, and the evaluator learns what "behavior" means when the artifact is a test suite.

File	Added	Removed
rules/harness-engineering.md	evaluator may read the subject to design a check, verdict from what ran; a test suite's behavior is failing on a broken subject; counts / deletions / "no longer applies" are load-bearing and must be enumerated before/after	/goal line; ADR format details (kept a one-line rule); clause duplicated in workflow
rules/workflow.md	"never claim a fact without proof" widened from completion claims to any stated fact (state what the output shows, not what you inferred); detectors need a positive completion signal; existing worktrees must be clean before dispatch	114K-token anecdote

Bullets 159 → 158. Still above the 150 budget noted in harness-meta.md; deletion candidates in development-principles.md are a separate decision.

Why these sentences and not others

Ran empirical-prompt-tuning with fresh executors on small fixtures reproducing 8 real misses from one long session. Only one reproduced without the new text: when evaluating an implementer's fold of 12 tests into a table, the executor verified the one disputed test by mutation but accepted the other 11 on the report's categorization. That is the sentence that landed ("a summary that adds up is not evidence").

Three candidate sentences for development-principles.md (enumerate branch-exclusive inputs before folding, convention over majority, validate CLI input before destructive use) were not added: fresh executors already did all three without them. Those misses came from a long main session evaluating its own work, not from missing text.

Evidence

Scenario	Red (current rules)	iter 1	iter 2	hold-out
A evaluate a fold report	✗ 70%	✓ 100%	✓ 100%
C mutation-detection harness	✓ 90%	✓ 90%	✓ 100%
B cluster names in a PR body (sealed)				✓ 100%

Judges were separate sonnet agents that saw only the deliverable and the checklist. Residual unclear points (what counts as "one condition"; whether merging alone triggers enumeration) are recorded in the log and accepted as definitional residue.

Eval set: docs/evals/2026-09-15-verification-claims/ (outside .chezmoiroot, so not deployed to ~/.claude) (fixtures, frozen checklists, per-iteration log, v1/v2 candidate texts). Fixture names are neutral.

Not in this PR

• chezmoi apply of settings.json — the live file has drifted from source (proxy env vs. deny list); apply the two rule files selectively.
• A hook for multi-line rationale comments in code (the existing rot-comment hook only matches English issue/temporal patterns).

🤖 Generated with Claude Code

[changeset-bot]

⚠️ No Changeset found

Latest commit: fe81dea

Merging this PR will not cause a version bump for any packages. If these changes should not result in a new version, you're good to go. If these changes should result in a version bump, you need to add a changeset.

Click here to learn what changesets are, and how to add one.

Click here if you're a maintainer who wants to add a changeset to this PR