[Brian Krabach (bkrabach)]

Busy manager conversations repeatedly scanned retained coordination receipts while constructing the latest request view. This change selects at most 32 IDs from each existing sender/target index before fetching the latest 32 exact receipts. Insertion order, self-send deduplication, transaction visibility and cancellation projection are preserved.

Release 0.96.4 is a patch for existing performance behavior. Its authored release entry needs no high-impact notice: permissions, defaults and recovery behavior are unchanged. Main owns the single normal publisher and authorized adoption, Spark-2 first, then Spark-1 and MacBook.

Validation:

• 143 distinct focused checks pass: five new performance/semantic checks and 138 existing coordination, generation, cross-workspace, cold notification and cancellation cases. The predecessor fails the three new cost checks and passes the two semantic checks.
• Actual AppService fixture, 25,000 retained 2 KiB requests: warm current-read CPU median 69.247 → 0.069 ms; SQLite work 319,708 → 1,955 instructions. Candidate work remains 1,955 at 50/5,000/25,000 rows. This is synthetic query scaling, not a live end-to-end speed claim; the small-case VM count increases from 1,603.
• Independent exact source/evidence review clear. Normal owned-DTU release prepare/build and locked dependency dry-run pass. Built client assets remain byte-identical to 0.96.3; client source, dependency selections and all qualified feature inputs are unchanged by the release metadata commit.

Grants, proposals, watches, initial index creation and other host latency remain outside this query bound. Required CI and published-package smoke retain their normal scopes.