[colosair]

Summary

Docs and tests only. packages/*/src is byte-identical, and so are the
manifests, the schema and the lockfile — the published 1.7.0 artifact is not
affected by anything here.

Two things the 1.7.0 release note said were true were not, and two behaviours
it relied on were not pinned anywhere. This corrects the first pair and
covers the second.

Changed

docs/releases/v1.7.0.md — ## Known limitations, both paragraphs.

JAM_WRITE_UNCERTAIN was recorded as having had "only its code path read".
It has had unit coverage since before this release (write-apply.test.ts,
plus create-apply, assignee and write-adapter cases). Left standing, the note
sends the next person to rebuild tests that exist. What was genuinely not
done is producing an ambiguous transport failure against real Jira, and the
paragraph now says that.

The --input paragraph blamed quoting. Measuring puts the cause in
PowerShell 5.1's native argument passing, and narrows the claim:

Git Bash        '{"text":"..."}'          works, spaces or not
cmd.exe         "{\"text\":\"...\"}"      works
PowerShell 5.1  value without spaces      works in some forms
PowerShell 5.1  value with spaces         failed in every form tried
PowerShell 7+   not installed here        unmeasured

The note now presents this as what a handful of forms did, not as proof that
no quoting can work. The --input-file / stdin follow-up note stays — it was
deferred this round, not dropped.

packages/server/tests/unit/write-transport-parity.test.ts (new).

The same refused request driven through the MCP tool and through
runJiraWrite, over one deps, with the codes compared. Running both is the
evidence; the source check underneath explains why they agree and catches a
second operation list being introduced, which is what would make them start
to diverge.

packages/server/tests/unit/write-apply.test.ts — one case added to the
existing a write whose outcome is unknown block. After UNCERTAIN, a
retry with the same planId gets JAM_WRITE_PLAN_NOT_FOUND and never enters
the write port. Provable rather than asserted: the armed failure is spent by
the first attempt, so a second call would have succeeded and left a comment —
none is there.

Review points

1. The fourth parity case does not match, and is recorded as it measures.
   operation: "issue.delete" is turned away by the tool's z.enum before
   the application sees it, and named JAM_WRITE_OPERATION_NOT_ALLOWED by
   the application on the shell path. Both refuse and neither reaches Jira,
   which is the invariant worth holding; asserting one shared code there
   would assert something untrue. The plan for this change expected a match —
   running it is what corrected that.

2. The UNCERTAIN retry case proves an absence by consuming the arming.
   FakeJiraWrite.failNext clears itself after throwing, so a second entry
   into addComment would have recorded a comment. No comment plus an
   unchanged mutation count is what says the port was never entered. It
   avoids pinning a harness counter as the contract.

3. What the completion gate is. No single line proves behaviour is
   unchanged — production diff 0 is one input among several, since a
   test-only change can still move packaging or a CI gate:

packages/*/src diff                      0
package.json / config / schema / lock    0
new command, flag or tool                0
new error code, state or dependency      0
npm test                                 544 pass / 0 fail / 4 skipped (548)
npm run release:check                    consistent at 1.7.0
tsc --noEmit (packages/server)           green

Tested

• npm test — 544 pass, 0 fail, 4 skipped. Was 538 pass; the 6 added are
  5 parity cases and 1 retry case.
• npm run release:check — consistent at 1.7.0. The version pins in the
  note were not touched.
• npx tsc --noEmit -p packages/server/tsconfig.json — green. Note that
  this config excludes tests/, as it does for every existing suite, so
  it covers src — which is unchanged.
• npm run release:verify — not run. It packs tarballs, and no manifest
  or source input to a tarball changed. CI runs build and test on both
  platforms.
• Published-artifact acceptance — deliberately not re-run. Nothing here
  reaches a published artifact; the gate above is what judges that.

Not in this change

--input-file and stdin, plan management commands, any write shortcut, and
transport auto-fallback. Each would add a step, a flag or an error for
someone who did not ask for one.

🤖 Generated with Claude Code