fix(update-change): draft the requested edit in step 4, write only in step 5 (#1840)

* fix(update-change): draft the requested edit in step 4, write only in step 5

Step 4 told the agent to "Apply the requested edit" while step 5 and the
guardrails told it to write only after the user confirms each revision.
"Apply" is a write verb in this very document - step 5 is titled "Confirm
and apply" - so the same request either wrote immediately or stopped and
showed the revision first, depending on which passage the agent weighed.
Step 5 is the workflow's only write path, so its confirmation guarantee
was unenforceable whenever step 4 governed.

Step 4 now drafts and says explicitly that it writes nothing; step 5
claims every write and shows the drafted edit for confirmation. Both
delivery surfaces and the committed skill carry the same wording, and a
regression test slices step 4 out of each body so no write verb can
reappear there.

Closes #1836

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

* fix(update-change): keep every edit verb out of the write-free step

Adversarial review of the first commit found three gaps.

Step 4 still opened a bullet with "Revise only files that already
exist" - the same shape as the bug, an imperative edit verb inside the
step that now declares it writes nothing. It reads as a scoping rule,
but "revise" is the write verb everywhere else in this body ("proposed
revision", "Which artifacts were revised"). It now says "Propose
revisions only to files that already exist".

The guard's `not.toMatch(/\bWrite\b/)` was inert and inverted: it was
case-sensitive, so it never matched the wording it was meant to pin,
and it could not be made case-insensitive because the fix's own text
says "do not write anything yet". It now strips that one sanctioned
sentence and rejects any remaining form of write or apply,
case-insensitively - so lowercase "write the drafted edit now", the
dangerous case, is caught. A second assertion rejects any step 4 bullet
opening with Revise/Edit/Update/Rewrite.

The changeset claimed no write verb could reappear in step 4, which was
not what the old guard did. It now states what the guard checks.

Also passes the surface label into section() so a marker drift names the
surface that broke.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

* fix(update-change): let no passage outside step 5 authorize a write

Mutation-testing the first guard found it defeatable: 20 of 25 mutations
broke the #1836 contract and still passed. The two worst were structural,
not lexical - the guard only sliced steps 4 and 5, so an authorization
placed in the intro, in step 3, in the Guardrails or in the Output
section governed the agent while no assertion ever saw it; and step 5
carried only positive assertions, so its gate could be kept and then
exempted in the next sentence, or the whole-body confirmation guardrail
deleted outright, with nothing failing.

The guard now pins step 4's draft rule and the whole of step 5 verbatim,
requires the whole-body guardrail to survive, and scans every other
passage for write verbs and their synonyms (commit/save/persist/
overwrite/reapply/flush/emit), verb-free equivalents (perform, carry
out, in place, to disk), consent-bypass phrasing, and imperative edit
bullets including ones led by an adverb. All 22 mutations are now caught.

Two wording corrections came out of the same review. "nothing earlier
writes to disk" was false - every openspec invocation persists a
telemetry id via the root preAction hook - so step 5 now claims every
artifact write instead. Step 4 says "in the conversation, not in files",
borrowing explore.ts's phrasing, so "draft" cannot be read as writing a
draft file.

docs/commands.md carried the same apply-vs-confirm collision two lines
above the confirmation bullet, contradicting the worked example directly
below it; it now says "Drafts your requested revision".

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

* test(update-change): catch more imperative edit verbs outside step 5

CodeRabbit noted `- Modify the artifact now` slipped past the
imperative-bullet guard. Added Modify, Amend, Patch and Replace, each
verified to trip the guard. `Change` is deliberately excluded: step 6
already opens a bullet with "Change already implemented ...".

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

* test(update-change): prove the write-gate scan trips in every section

The outside-step-5 scan was only ever checked against the real body, so
nothing showed it could fail. Injecting a verb-free authorization into the
intro, Input, steps 1-2, Output or Guardrails passed on both surfaces (7 of
10 sections). The bare "already applied" allowlist entry also erased
"treat the requested edit as already applied" before any check ran.

- Move the scan into a function and add a mutation table: one injected
  authorization per section, asserted on skill and command bodies.
- Spell every sanctioned mention in full context.
- Flag any mention of the requested edit outside the pinned draft rule.
- Widen the consent-bypass filter (needs no / exempt from / skip confirm).

Also revert the legacy docs/commands.md edit: docs-lab reference/skills.md
is the published page and already states the confirm-then-write contract.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

* docs(changeset): drop em dashes from the release note

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5 <noreply@anthropic.com>