[helly25]

Reduce overhead when a persistent matcher pool handles a short tail after a larger batch. Keep useful read concurrency, existing worker creation and output ordering.

AG;DR

• Keep four-entry dynamic chunks and current serial/creation admission. A batch activates at most one retained worker per chunk; every thread remains available for later batches.
• Give workers stable IDs and keep eligibility fixed throughout each generation. Only participants decrement completion; stop wakes all retained workers, including inactive ones.
• Exercise growth, 64-entry batches retaining twelve useful workers, 16/32-entry tails, empty batches, regrowth, inline tails and destruction with inactive workers. The regression records actual reader thread IDs.
• Add an explicit manual warmed-tail benchmark and a controlled delayed-read variant. Reject the alternative that applies the creation quota to each batch: it wins cheap kernels but slows delayed reads about 95-294%.
• Measure clean ThinLTO release executables on this macOS ARM64 machine and Linux x86-64 drogon. Each host has 49 native cases x three sessions, 18 in-memory cells x three sessions, and four controlled-latency cells x three sessions. Sampling retains the fastest seven of nine, with rotating participant order and native output oracles. Linux timing pins CPU allowances; macOS has no affinity guarantee.
• The selected 16-entry/ten-worker warmed kernel improves about 31% Linux / 51% macOS. Delayed-read concurrency stays effectively unchanged. Native gains are modest and some controls are slower: retain five-session confirmations with a duplicate-baseline participant, including Linux filter costs up to about 4.13%. Do not project the kernel improvement onto whole-command throughput.
• Preserve initial/rejected/invalid-delay experiments and exact source/binary identities in the lossless evidence archive; explain the fixture correction explicitly. Link the report from the existing O11/JC checklist and release site. O12 adapter simplification and stage scheduling remain separate.

Validation: complete core/extension release suites pass 306/306 on both hosts, with required Linux FUSE mounts; focused macOS ASan and Linux TSan each pass three suites. Six affected translation units pass local scoped clang-tidy. Normal enforcing hooks and pushed-head CI are required before merge.

Based on merged #1050 (2d615ea1e447e55c266a47799ad6cd64bd4e246d). The synchronization has exactly the locally verified tree; the recorded measurement parent has exactly the merged #1050 tree. Current pushed-head CI remains authoritative.