[helly25]

Reuse matcher output buffers within bounded retained capacities instead of reallocating them for every batch. A repeated 64-entry output batch uses 20% fewer allocation calls and about 25% fewer requested bytes on both macOS and Linux, while preserving output and error behavior.

AG;DR

Result ownership and reset contract

• Reuse the existing ParallelResult slots. Workers mutate only their exclusively claimed private slot; the private entry evaluator receives its index and writes the pool's own state directly.
• Before any new evaluation, release old text buffers whose capacity exceeds min(64 KiB, 1 MiB / current batch size). Shrinking/empty batches destroy truncated tails.
• Clear reused text, reset selection status, and disengage content when a retained matched slot becomes unmatched. Preserve false/error/true transitions without stale output.
• These limits apply to text capacity retained from a consumed batch. They do not truncate current output or cap status messages, vector metadata, or total process memory. Current records remain pool-owned and available until the next batch or destruction.
• Keep input consumption, ordered result delivery, generation barriers, worker counts and admission policy. No input-recycling API, extra owner or new CLI flag is added.

Qualification and rejected controls

The committed decision report at docs/benchmarks/bounded-match-results-20261102.md and its lossless archive retain all source variants, source/executable hashes, fixture definitions, raw repetitions, scripts, allocation counters, and verification logs.

Four variants separate the mechanisms: baseline, direct writes while clearing/recreating records, bounded matcher reuse, and comparison-vector retention only. Adopt bounded matcher reuse. Reject the reset-only change as a separate optimization because it saves no output allocations. Reject comparison-vector retention: VFS range reads still return fresh owning strings, and keeping old result slots can prolong old small patch payloads without reusing their input buffers.

• Coordinator-only ASan counters for a warmed 64-entry/1,028-byte-per-line batch: 320 -> 256 allocation calls on both hosts; requested bytes 264,704 -> 198,656 on macOS and 266,752 -> 200,192 on Linux. Setup, inputs, warmup and assertions are outside recording. These are allocation traffic measurements, never sanitizer performance timings or net live memory.
• Native timing: 61 cases x three sessions x four variants, discarded warmups, nine rotation-only rounds/fastest seven, independent output oracles. Linux uses tmpfs and 1/3/10-CPU affinity; macOS uses warm APFS without affinity enforcement.
• Separate case-interleaved in-memory timing: 46 RE2 cells plus 16 comparison cells, three sessions, nine repetitions/fastest seven. Linux pins 1/4 CPUs; macOS uses worker allowances. Eight focused cases repeat both promising and slower/noisy controls.
• Preserve slower samples. Original macOS large-output ten-file timing is +4.90%; focused confirmation is -1.42%. The macOS 1,000-file/four-worker filter control remains +1.66% in confirmation. Native worst selected medians are +2.69% on Linux and +1.63% on macOS.
• The comparison-only control also changes unrelated matcher timings. Their cause is not established; do not attribute those gains to output-buffer reuse. Copied benchmark executables have no mutable project shared-library dependency.
• Unstripped build outputs shrink by 16 bytes on macOS and 2,328 bytes on Linux. No universal throughput or total-memory improvement is claimed.

Validation

• Complete core/extension release suites pass 306/306 on both hosts; Linux uses required-FUSE integration.
• Focused macOS ASan, Linux TSan, and final retained-slot regression release/ASan tests on both hosts pass. The final Linux regression also passes TSan.
• Scoped local clang-tidy covers all five affected translation units, with the test-only arithmetic finding corrected and rechecked. Normal commit/push hooks enforce the current files; pushed-head GitHub CI remains required before merge.
• Tests cover small-buffer reuse, aggregate and per-record reuse limits, giant output followed by a short inline tail, a retained matched slot becoming unmatched, empty-batch/growth transitions, and false/error/true status transitions without stale text.
• Native wall/CPU time and child peak RSS remain in the evidence. First-output latency is not separately measured. Builds/lint/sanitizers never overlap timing on the same host.

Update TODO.md, docs/performance-analysis.md, CHANGELOG.md, and the release-site report mapping together. The exact evidence archive gets a narrow large-file-hook exception alongside the existing reports.

Based on merged #1049. The measured parent 8d96bc90fea5adb469b42e5e34a8be6a518f0f4b and merged main 901c8fea916124f4d0e195c5385f2a977b44093d have identical trees; the unpublished branch was synchronized before committing. O11 active-worker participation and budget-aware comparison source reuse remain separate.