fix: a stamp the notary reports written is a stamp that reads back — and 3.3.0

[TinDang97]

One defect fix, then the 3.3.0 version bump.

The fix — a stamp the notary reports written is a stamp that reads back

_oneline exists because an unbalanced { in a --reason once made the parser swallow the following stamp — two records written, one read back, from an append-only ledger whose ordering is the trust model. That fix was correct and was applied to exactly one field. Seven writers interpolated by: raw.

$ add freeze t --by 'O"Brien' --authority human
freeze recorded at authority `human`     # reported as SUCCESS

stamp keys read back: ['by']             # act, authority, direction swallowed
_is_frozen       -> False
sealed_direction -> None

The seal silently does not exist while the human is told it does. The trigger is an odd number of " — a balanced pair (Tin "TinDang97" Dang) round-trips, which is exactly why this survived every real use of the engine.

Severity, stated plainly: it fails CLOSED. The gate then refuses with R:UNSEALED, so nothing is let through. The defect is that a notary whose only job is to record faithfully reported a record it did not write → R:LIE.

All seven interpolations (freeze · brief · replan · check · interview · both gate paths) now go through _oneline, and the reason is written at the function's own definition, since the fix is one careless edit away from reverting. test_every_stamp_writer_normalises_its_by enumerates the writers from the source, never from a hand list — a hand list is precisely how the seventh writer got missed, which is the shape of the defect itself.

The gate refused the first PASS: M4 ("a verb reports a refusal, never a false success") was declared with no check. Rather than staple a covers: onto an existing one, the punctuation walk now asserts the biconditional — what the verb reports must match what the ledger holds. A success that wrote an unreadable record and a refusal that wrote a good one are both the notary lying.

Corrections to my own earlier report, disclosed rather than quietly fixed

• I first reported this as a brace bug that destroyed the entire verified: ledger. That was wrong. The probe behind it froze an unauthored node, so freeze correctly refused on template placeholders and wrote nothing — I read a correct refusal as corruption. Corrected on feat: RISK-ACCEPTED signs for weak evidence never a missing seal, and the ONE approval asks its questions out loud #208 as well.
• M1 (_oneline neutralises the quote) was already true before this task; _oneline has always replaced ". It stays as a regression pin, but the real rule is M2 — the writers that never called it.
• test_no_existing_stamp_is_rewritten first compared whole-file text, which a later write legitimately grows. It compares the parsed stamp list now.

The release — 3.3.0

Six version sources in lockstep: pyproject.toml · package.json · package-lock.json (2 places) · .claude-plugin/plugin.json · src/add_method/__init__.py · tooling/add.py's ENGINE.

Bumping ENGINE re-aims add.py, so the four engine twins are resynced, ENGINE_MD5 re-aimed, and both dogfood bundles' engine:/tooling_engine: stamps moved — those ride every version bump and go stale silently otherwise.

The CHANGELOG covers what #207, #208 and this PR shipped: the scaffold beat · the claimed-output guard · the INTEGRITY/EVIDENCE split · done's seal check · the template-Milestone refusal · segment-aware _paths_touch · sensitivity flooring up · add interview · this stamp fix. It says the RISK-ACCEPTED hole failed closed and that the stamp defect made a verb report a record it had not written, because a changelog that only lists wins is the same failure mode as a guard that only checks the happy path.

873 passed / 0 failed / 7 skipped across both roots.

Not in this PR

Tag and publish remain the human call. Pushing v3.3.0 triggers the npm + PyPI publish, and neither registry lets a version be reused. Nothing here pushes a tag.