[helly25]

Keep archive insertion unchanged. Retain two measured insertion experiments and their counterexamples so future work can avoid repeating them, and close stale O30 tracking for the owned-input optimization already merged in #1122.

AG;DR

Decision and retained scope

• Documentation, evidence and tracking only. No production C++, benchmark implementation, CLI behavior, worker policy, build default or official historical measurement changes.
• O27a deferred ancestor-key ownership previously completed 28 public construction, 82 command and 88 listing cases on both hosts, with three balanced sessions, nine repetitions retaining seven and an unchanged-parent control. Its modest lifecycle gains did not establish an attributable cross-platform whole-command benefit. Preserve the complete original archives byte for byte; do not rerun the duplicate mechanism unchanged.
• O27b direct mapped-node construction is a distinct screen on main 9bf1aad510d843d1d6338fbdf23107a30d339730 (Record shared-worker qualification and retain specialized pools #1125). Both hosts passed all 28 public construction fixture oracles and three focused optimized archive targets. The one-session screen uses three repetitions retaining two, with baseline/candidate/identical-baseline participants. All Linux shape medians are slower (+0.60% to +2.34%); apparent Mac gains largely track its unchanged-image control. Large tiny-case disagreement and every slower result remain in the evidence. Stop at that screen: no new full-root suite, sanitizer qualification or full three-session population is claimed.
• Keep transparent lower_bound plus hinted ancestor insertion as separate, unmeasured O27c. Preserve duplicate replacement, synthesized-directory state, direct-child counts, inode order and source ownership.

Evidence and presentation

• Add docs/benchmarks/archive-index-insertion-20261107.md and its lossless 4,525,832-byte .json.gz capsule, SHA-256 ef0f035c7369a11f8191f9b46407b6682cfb93ce1190e70743172fd0ed360527.
• Retain original contains-first archives, the new source patch, source/executable identities, every screen observation, drivers and build/test logs. Different parents/protocols remain separate populations.
• Register the report in release-site.json; extend the large-file exception for this exact archive only. Positive and negative filename controls reject unrelated files, dates and suffixes.
• Split O27 tracking into the two completed rejection decisions and the remaining hinted-insertion candidate.

O30 tracking correction

Owned OpenBytes input transfer already merged in #1122 as 0e95e191ba7d7ee3769217e7fbf82a825ff9b90d. Its exact push/main Test run 42276661524 completed successfully, including both operating systems, sanitizers, coverage, GCC, fuzzing, CI clang-tidy and policy checks. Remove the stale create-a-PR instruction; retain O30's original noisy Mac measurements and private-versus-public image limitations.

Validation and dependencies

The exact six-file payload was copied with per-file SHA-256 verification into an independent Linux checkout at the current main base. Changed-file pre-commit run --files ... passed with SKIP=clang-tidy as explicitly requested by the user; no local clang-tidy was run. All 32 existing release-site tests, configured report mapping, precise allowlist controls, gzip/JSON decoding and git diff --check passed. Every payload hash remained unchanged after these checks. The release-site unit tests use their existing local renderer fixtures; no unpublished-report API preview was attempted.

Normal signed commit 11926909998d762a29d9e32b972455b71cf10770 and normal push completed; the signature verifies as good. This private clone has no automatic commit or push hooks configured, so no additional hook execution is claimed. The independent Linux changed-file checks above passed on the exact published bytes. This PR still requires its own current-head CI, separate from archived experiment qualification. Based directly on merged #1125/main; no open PR dependency, and no change to the preserved unqualified #992/#996/#997/#998 branches.