[helly25]

Reuse each file's validated rg output as the matcher result instead of replaying and copying it. This reduces allocation traffic and improves large-output workloads while preserving native callback behavior, late-error handling and binary suppression. Ordinary whole-command search controls remain effectively unchanged.

AG;DR

Implementation and boundaries

• RenderRgOutput returns the owned per-file output transaction and selection state. ParallelMatch transfers its existing reusable result allocation into rendering and takes the validated transaction back, avoiding the outer replay/copy and unused rg record-end storage.
• Specialize only the ordinary rg publication path. Native grep retains its original local buffer, callback boundaries, frozen sink and bool/status return. Reduced and multiline rg keep validated capture. Restore heading and sink state on both success and failure; discard all output after late read failure or a binary marker.
• Preserve the existing retained-capacity limit, worker counts, batching, CLI behavior, source ownership and serial archive barriers. This is not output streaming, pipeline overlap, a global byte bound or an oversized-output solution; O18 remains open.
• Add committed regression cases for native callback boundaries/context separators, sink replacement, late read and binary failures, heading/state restoration, nonmatching scratch and reused-result capacity. The ASan fixture measures allocation counts only, never performance.

Evidence and limits

• The report and lossless evidence are docs/benchmarks/transactional-rg-records-20261107.md and .json.gz (5,079,513 bytes; SHA-256 f32ec19cd3af284576a8f6af7e84fbedbf10dee48b044bfdc05909e0ffc7c466). The archive preserves all three variants, counterexamples, copied-image controls and complete source/raw/log records; no experimental runtime variant other than the selected native-in-place form enters this PR.
• The selected 64-entry allocation diagnostic removes three allocations per entry with unchanged retained result capacity. Memory large-output cohort times improve 12.36% on Mac and 4.32% on Linux. Ordinary native rg-line cohort changes are -0.23%/+0.00%, with same-binary repeats -0.31%/-0.34%. Individual large native-output cases improve up to 11.35%/50.56%; these are not general CLI claims.
• Retain costs: Linux native-line memory cohort +3.18%, Linux PCRE2 count/100/4 +8.53%, Mac PCRE2 native-lines/10,000/4 +6.63%, and all slower native controls. Do not attribute every difference to allocation traffic, capacity or unrelated-system noise.
• The experiment measures lean release images based on c95140674254c371844f1fab27a64951a80dac55, not sanitizer timing or current release distributions. Lean images grow +2,816 bytes on Mac and +11,240 on Linux. Mac timings are unpinned; Linux uses pinned physical-core-separated logical CPUs. No native RSS improvement is claimed.
• All retained variants comprise 101,736 raw measured observations, excluding warmups: three variants, both hosts, complete 128-case memory and 124-case native populations with three balanced sessions and nine repetitions retaining seven. This is not a selected-only population or total CLI-invocation count.

Documentation and publication

• Update CHANGELOG, performance analysis and O22c2h tracking; keep O18's byte admission, ordered completion, head credit, cancellation/draining and oversized transaction requirements explicit.
• Register the Markdown report and supported release-site files mapping for the exact gzip download. Add only its precise large-file exception, preserving all existing report/history entries. This adds about 5.08 MB per rendered release snapshot; no historical performance data is removed.

Validation and dependencies

The isolated development candidate passed all 323 documented core/registered-extension release targets on both hosts, including required real Linux FUSE, plus four focused canonical ASan targets per host. Source/evidence identities and all record hashes are preserved independently of public PR readiness. No local clang-tidy is run, as explicitly requested by the user.

Current integration is based on merged #1126/main d8cf61cccd4b0efed5058bf6fa45228e47162b48; five C++ source/test files and the frozen report/evidence were preserved byte for byte while adding publication metadata. Current-base integration passed the three optimized semantic targets (rg_test, parallel_match_test, evaluate_test); these cached results were accepted only after exact input verification. The exact twelve-file changed-file pre-commit gate passed with SKIP=clang-tidy; all 32 release-site unit tests, supported files/page configuration, precise allowlist positive/negative controls and diff checks passed without altering the staged bytes. O22c2h tracking is marked integrated; its own PR CI remains required. Normal signed commit c4d21aea2f6640386d49c4a2f66503472bd64f96 and normal push succeeded; its GPG signature verifies as good. At the initial publication this private clone had no automatic commit/push hooks configured, so the exact manual changed-file gate above is the initial local hook validation claimed. The later scoped Trunk formatter initialized its configured Git hooks; the subsequent signed Markdown correction and push ran those hooks successfully. No hook was disabled or bypassed. This PR must pass its own exact pushed-head GCC, coverage, sanitizers, CI clang-tidy, final benchmark aggregate and required checks; archived experiment results do not replace them. The four unqualified historical worker PRs #992/#996/#997/#998 remain untouched and are not dependencies.

CI Trunk identified one extra blank line between performance-analysis sections on the initial head. Follow-up signed commit 92469874d258200a1e07fcbabcc9c8b150602294 removes that single blank line; scoped trunk fmt recheck and changed-file pre-commit passed. All C++ source, tests, report and compressed evidence bytes remain unchanged. Healthy reachable collectors were not cancelled; the updated head requires its own CI.

CI clang-tidy then identified three narrow source issues. The follow-up keeps the selected native buffer and record-boundary vector local, makes the unused non-publishing specialization's boundary vector const, accumulates already-computed boolean matches with |=, and removes redundant typed-lambda parentheses. Fresh Mac development checks passed rg_test, parallel_match_test, and expression_allocations_test (three executed tests); scoped formatting, changed-file hooks with explicit local clang-tidy skip, and normal signed commit/push hooks passed. The immutable experiment archive still describes the original measured source; no byte-identical performance claim is made for this style correction. Fresh pushed-head CI, including actual CI clang-tidy, remains authoritative. No healthy collector was cancelled.