Pre-#2448 cursor[bot] driver ledger stands as an unsupersedable not-clean verdict (blocks #2341) #2482

The state

PR #2341 is blocked by check-pr-fully-clean.py's per-reviewer latest-verdict rule (#2274):

• cursor[bot] (a Cursor cloud driving session, not a reviewer) posted a disposition ledger on 2026-08-26 ending "Do not merge. Blocked on review of 8af4edc" with an unbackticked "(Needs more work)" quote in its first line. The checker admits it as Cursor's standing not-clean verdict.
• The checker keys the Cursor identity EXCLUSIVELY to the cursor login (EXCLUSIVE_BOT_IDENTITY), so only a comment posted by Cursor's GitHub App can supersede it.
• That Cursor session is gone (claim expired), and no local process can post as cursor[bot]. cursor-agent CLI reviews post under the user's own login and are bucketed separately.

So the PR carries a permanently standing not-clean verdict that no reachable process can supersede: an undischargeable state, blocking mwc merge forever, however many clean Claude rounds land.

Why #2448 does not cover it

#2409 (driver ledgers admitted as verdicts) was closed by #2448's convention: backtick quoted verdict phrases in ARD disposition comments. That is forward-looking --- it fixes comments written after the convention. The 2026-08-26 comment predates it, and its author is not around to edit it.

Fix directions (pick one)

1. Time-box exclusive-login verdicts the way claims are time-boxed: a not-clean statement from an exclusive-login bot whose session's claim has expired (>2h no push/comment from that login on the thread) and which is phrased as a driver ledger ("Do not merge. Blocked on review of X" + a Disposition table) stops standing once a review OF THE CONTENT IT NAMED lands clean.
2. Manual dismissal affordance: teach the checker to honor an explicit human dismissal comment ("dismissing cursor's 2026-08-26 statement: driver ledger, condition satisfied") from the repo owner.
3. Retroactive classification: apply the ard: backtick a quoted verdict phrase, instead of teaching the checker to guess #2448 backtick test only to comments after the convention's merge date, and classify pre-convention comments that match the driver-ledger shape (disposition table + hold language) as ledgers, not verdicts --- accepting check-pr-fully-clean.py: driver-comment classifier fails open on a Copilot-format review carrying a real finding #2430's fail-open concern is why this needs a POSITIVE signature (the cursor login itself + the table + the "Blocked on review of " form).

Until fixed, #2341 (and any PR carrying a pre-#2448 cursor[bot] ledger) needs human action to merge.

This issue body was drafted by Claude Code (AI agent).