[helly25]

Avoid unnecessary temporary strings when recording collection names and comparison categories while preserving owned keys. Retain the complete cross-platform executor/copy audit, including useful memory gains and counterexamples, so future optimization work can build on reproducible evidence.

AG;DR

The implementation retains only C01/C03 from the copy audit:

• Collections::Add uses transparent try_emplace(name) instead of constructing std::string(name) for every lookup. New keys remain owned; the regression test mutates/destroys caller storage and checks repeated-name lookup and retained entries.
• Comparison category recording accepts string_view and constructs an owned map key only when retention is needed, avoiding the eager std::string(pair.path) temporary.
• The two production C++ files are byte-identical to the measured a49c17045d candidate. The regression test has matcher-only EqualsText style corrections and the direct matcher dependency; the immutable archive retains the original measured source/test patches. No recycler API, executor integration, admission gate or scheduling behavior is added.

The new report and condensed performance-analysis entry cover all seven original PRs (#991, #992, #996, #997, #998, #1000, #1001), actual dependencies, isolated/cumulative variants and eager mechanism controls. Keep main's scheduling: useful Linux gains coexist with macOS collection and specific content regressions. Removing the inherited 8192-entry/500us matching guard repairs the largest content regressions on both platforms, without establishing a universal shared-pool win. The bounded frontier has a demonstrated case-specific memory benefit: in the 65,536-file metadata case at three workers, #997 reduces median peak RSS from 20.06 to 6.25 MiB on Linux and 17.34 to 5.44 MiB on macOS (about 69% vs #996); this is not a worst-case bound. C02 recycling remains deferred. C01/C03 are narrow construction/ownership fixes, with no universal wall-clock promise.

The lossless docs/benchmarks/executor-copy-audit-20261101.json.gz archive contains ten complete report populations, fifteen full binary-capable source patches, the earlier #991 reassessment, raw samples, drivers/fixtures and provenance corrections. Its exact size is 4,351,250 bytes; SHA256 is b10a92105d69154e933ccfa4f5bb1776ad27e145c80e9b9f8de976da259c7560. The initial native matrix has 53 cases, three sessions and 15,900 oracle-correct invocations per platform, followed by separate native/in-memory controls. The report preserves CPU/storage differences, original manifest labels with explicit tree-equivalence correction, full source identities and negative controls. Collector Google Benchmark threads are distinguished from engine job budgets; its short-key/checksum contract is weaker than native complete-output oracles.

Register the report in release-site.json; add the allocation note to CHANGELOG and remaining qualification work to TODO. The archive exceeds the normal 500 KB pre-commit limit, so check-added-large-files has one exact-filename exception for this evidence artifact; other hook checks remain enabled.

Validation: current pushed head 542ca47120e8ded422e88fc9de66972f45e7ceb0 includes current main 8ff28cde9c67bd498f9016935da01506ffb5845c. CI run 41307562869 completed SUCCESS; all 32 reported contexts are successful or skipped, including both default platform suites, ASan/TSan/MSan, GCC/minimal, clang-tidy, coverage/fuzz, release-site, six benchmark measurement shards, cross-platform comparison and final done. Normal commit/push hooks passed; the synchronization push ran actual clang-tidy on all five inherited infrastructure translation units after a scoped compile-database refresh. Focused collection/matcher/rg/run tests previously passed on both platforms for the measured copy candidates, and the matcher-only test correction received fresh focused collection validation. The six-translation-unit audit lint covered retained C01/C03 code in the recycler superset rather than a separate a49-only lint run. Raw archive populations/source patches and error-free retained results were independently checked. These measurements do not validate unmodified runtime PR heads.

MSan timeout correction: //xff/cli:rg_test aggregates 33 functional cases with repeated frontend subprocesses. An earlier run passed 297/298 tests but this test hit its implicit 60-second small-test timeout, without a sanitizer diagnostic or failed assertion; an earlier successful infrastructure run took 55.3 seconds. Set only its timeout to moderate (300 seconds), preserving size small and every assertion/case. No application code, skipping, splitting or sanitizer suppression is introduced. Focused macOS release rg_test passed in 3.8 seconds; the corrected pre-sync Linux MSan run passed all 299 tests, with rg_test taking 68.5 seconds, exceeding the old limit. The final synchronized head also passes full MSan in its own CI context.

Portability/limits: release measurements use Clang 23.1.4/ThinLTO, Linux KVM/tmpfs with CPU affinity and native Apple M5 Pro/APFS without affinity. In-memory Google Benchmark results use a separate contract. Small timing changes and byte-identical negative controls do not establish causality or significance; physical slow storage, arbitrary worst-case RSS and full-plugin/PCRE2 startup are not established.

Dependencies: based on current main, including independently merged infrastructure #1035. That PR's synchronized head b830e17874 completed full CI run 41303081236 before the user merged it as 8ff28cde9c. Its test-only support has no shipping executor dependency. Post-merge main Test run 41307382327 completed SUCCESS, including default macOS and final done. The separate main Benchmarks workflow 41307382275 is in progress. These are separate contexts from this fully green PR.

Concurrent #1036 was independently classified as an isolated uploader inventory/display component: one read-only published batch/status scan plus set intersections outside the CLI/measurement loop. Independently merged #1038 adds uploader dataset-recipe overlap/incomplete-batch inventory and bounded publication rebase/regenerate/retry in Python/workflow/docs. Neither changes C++ runtime, benchmark measurement execution or retained raw samples; no additional timing is needed for these runtime claims. The immutable audit baseline remains historical main 974d137026, preserving measured identities rather than relabeling results after synchronization. Original runtime PR heads remain unchanged.

Ready for review only. No approval, merge or auto-merge is requested.