3.4.0 — a guard asks whether what a stamp attests is TRUE

[TinDang97]

Cuts 3.4.0. Two merged-but-unreleased PRs ship here: 3.3.0 was tagged before #210 landed, so this carries both #210 and #211.

What ships

#210 — the front door teaches a walk that runs. The documentation described a method the engine did not implement, and nothing could see the gap: every guard checked nouns the engine EXPOSES rather than capabilities the prose PROMISES. GETTING-STARTED.md is now executed by the suite, the roster names agents that exist, and init seeds selectable personas.

#211 — a guard asks whether what a stamp attests is TRUE. Every integrity guard asked whether a record was shaped right; none asked whether the thing it recorded had happened. A gate HARD-STOP walked to done. A red Task closed on another node's receipt through a slug collision. brief resolved no lens at all — not one seeded persona had ever reached a worker — and said nothing about the omission. And a scaffold new wrote could not survive the guards new recommended next.

Version surface

All eight declarations: five manifests (package.json, package-lock.json ×2, pyproject.toml, __init__.py, plugin.json), the ENGINE = "add/3.4.0" stamp, and the three shipped skill trees. Plus the dogfood bundle's own engine:/tooling_engine: in .add/index.md, which rides every bump.

ENGINE_MD5 re-aimed 9a800083 → 7932ec06. cli.py untouched, so ENGINE_PKG_MD5 stands.

One fix that is not in either PR

run learned to sniff the report path last turn, which left the shipped skill teaching the doubled --junitxml in four places — one of them a sentence explaining that the repetition was deliberate. Both spellings still work, which is why nothing failed.

test_receipt_idiom_truth could not catch it. It scopes itself to report-READING examples, decided by whether the engine-side flag is present, so dropping that flag makes an example invisible to the guard rather than wrong. That is the same guard-class hole this release exists to close, one turn later and self-inflicted.

The new guard binds the taught line to BEAT_NEXT["build"] — the string a user is actually shown at the end of a build — and was mutation-tested: reintroducing the doubling in verify.md turns it red, restoring turns it green.

Two smaller notes from that fix:

• My first cut of the guard counted --junitxml anywhere on a line and flagged my own corrected line for naming the override flag in a trailing comment. Narrowed to the argument text before #; a comment is not something a reader copies.
• The new prose collided with test_receipt_artifact.py, which parses --junitxml <token> and read my explanatory words (before, <path>) as documented receipt paths. I reworded the prose rather than touch that guard — it protects a real property.

SKILL.md holds at exactly 176/176 lines, funded by compression, no pin bump.

Verification

1111 passed, 7 skipped across both test roots (add-method/tests/ and add-method/tooling/), the full CI shape.

Tag and publish to npm + PyPI follow once this is green and merged.