Make changelog assembly recoverable after a partial write

`--write` wrote CHANGELOG.md and then removed fragments one by one, so a
failure after the write left surviving fragments that a rerun spliced in
a second time.

The write is now staged. The assembled file goes into
changelog.d/.assembling/, the fragments are renamed in beside it, and
renaming the assembled file over CHANGELOG.md is the single commit
point. A run that fails rolls itself back when it can; otherwise the next
`npm run changelog` reads the staging directory and either restores the
fragments (assembled file still staged) or deletes the leftovers (batch
landed). `--check` reports a staged batch instead of reporting past it.

A failure-injection test makes every filesystem call in a `--write` fail
in turn, both as a one-off error and as a crash with no in-process
rollback. It asserts the retry produces exactly the uninterrupted
CHANGELOG.md, with each entry once and an empty directory. The test
fails against the previous write-then-delete and against delete-first.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

changelog.d/README.md

@@ -38,6 +38,13 @@ The assembler refuses a batch rather than guessing: a filename that is not
 `<pull request>-<slug>.md`, or two fragments claiming one pull request number,
 stops the run before anything is written and names the files.
 
+If `npm run changelog` fails part-way (a full disk, a killed process), run it
+again. The batch is staged in `changelog.d/.assembling/` and only lands when the
+assembled file replaces `CHANGELOG.md` in a single rename, so the rerun either
+finishes that batch or puts its fragments back first. It never adds an entry
+twice or drops one. `npm run changelog:check` names a staged batch, if one is
+left over, instead of reporting past it.
+
 A documentation claim that needs a regression test — the way
 `src/annotations/drawTool.test.mjs` asserts against `CHANGELOG.md` — should
 assert against `docs/CURRENT-STATE.md` or `README.md` instead. A pull request