docs(watch): correct the three (#683) source comments (#988)

[fujibee]

Fixes #988. Follow-up to #987, which corrected the (#683) tags in tests/. The same miscitation survived in three scripts/watch.sh comments — and a source comment is read by the person about to change that code, so a wrong pointer there costs more than a wrong test name.

#683 is "Per-team migration copies a global read cursor without preserving its sequence floor" — read-cursor migration. Each comment was checked one at a time against that subject; keeping the tag was an allowed outcome (the goal is that every remaining #683 be correct, not that there be none). None of the three is about cursor migration:

line	comment subject	verdict
279	the sharing warning for two unfiltered watchers on unclaimed pairs	not migration, not the claim-race, no matching issue → tag removed
550	the held-pair set that de-duplicates departure/return announcements ("said" = announced, not read)	bookkeeping, not migration, no matching issue → tag removed
626	a stale process consuming a message for a pair another session claimed	the #983 hazard — this comment sits on the FIRST actas_lock_state check, and #983 is the still-open window BETWEEN that check and the later fetch. It names the hazard; it does not close it → retagged (#983)

Comments only — no code changes. bash -n clean, enforced-assertions at baseline. (#983 remains open; nothing here fixes it.)

Why one wrong number is worth this much care

The genealogy of this single miscitation — traced by git log -S to the introducing commits — is the whole argument:

one wrong (#683) on watcher-series commit subjects
  e3844c2 (subject did not even cite it; the tag was added to the body later)
  e4db061  "step aside while a pair is held … (#683)"      → CHANGELOG L275
  be48ecf  "release a pair another session has claimed (#683)" → CHANGELOG L276

lands in three kinds of place:

fixable	not fixable
tests/ tags — done (#987, landed)	git history: (#683) is in the squash-merge commit subjects, immutable
scripts/ comments — this PR	published CHANGELOG.md L274–276, regenerated by git-cliff from those subjects (a hand edit reverts next release)
	published release notes / distributed tarballs

Three people reached "L550 is not #683" by independent routes (the code de-dups announcements, not reads; no independent evidence it was ever judged #683; its tag descends straight from a miscited subject). The point of the asymmetry: a single wrong issue number is cheap to write and, once it rides a squash merge and a release, impossible to take back.