[droplister]

Summary

• process only plausible historical continuation transactions before the generic classifier-version backlog
• keep SQL as a queue hint; the raw-transaction classifier remains the authority
• bound concurrent R2/D1 reads to five and keep writes sequential
• run 50 reclassifications at the start of each recovery maintenance invocation
• persist completion of the one-time priority lane so later generic sweeps avoid repeated scans

Why

Classifier v2 is live and correct, but the generic queue contains about 501,000 transactions. Only roughly 35,000 transactions can contain the 2014 continuation outputs changed by v2. Prioritizing those makes the user-visible recovery set converge first without a bulk production rewrite.

Verification

• production EXPLAIN QUERY PLAN: classifier-version index plus exact primary-key join
• npm run check
• npm test -w xcp-api (499 tests: 465 pass, 34 skipped, 0 fail)
• Worker dry-run bundle: 752.8 KiB / 768 KiB