Fix Five Carried Skill Gaps a Downstream Review Raised

[ptr727]

Raised by CodeRabbit on ptr727/NxWitness#588, a resync carrying the hub's .github/skills at main 45669468. Each flagged text is byte-identical to its .agents/skills/ source here, unchanged on develop at 23ef3574, so the fix lands in the source rather than in the carrier.

• branching-and-release-model/references/release-publish-mechanics.md, "A full re-run is always available inside its window". GitHub also caps the re-runs of one workflow run (reported as 50, full and partial combined, not yet verified against the docs), so a run past that cap has no re-run route although the window is open. State the cap, or drop "always".
• drive-pr/SKILL.md, the comments label step. It requires the label when opening a pull request and notes that a label added later applies to the next push, but never says to add the label before a later push to an open pull request that adds or edits a comment line, so that push fails the prose gate. Related to A comments Label Applied by gh pr create Misses the Event That Starts the First Run #1703, which is about label timing on the opening event.
• merge-and-release/SKILL.md description, "this skill wins over branching-and-release-model, which supplies the policy it follows". "Wins" is meant as trigger routing, but it reads as the procedure overriding the policy. Say it wins the trigger while the policy stays authoritative.
• pr-review-conduct/SKILL.md outcome 2 says a decline happens "in the thread" and closes the thread, while a suppressed finding has no thread. The procedure exists in "Answering a suppressed finding", and outcome 2 does not point at it.
• unattended-handoff/SKILL.md picker eligibility. It checks for a backlog-burndown claim comment and then creates auto-<issue> as separate steps, and "one run per repository at a time" rules out two unattended runs but not an unattended run beside a backlog-burndown run. Both can pick the same issue and open competing pull requests. Needs one shared, atomic reservation.