[helly25]

Keep archive ancestor insertion unchanged. Record the final requested O27 mechanism and its complete counterexamples so future work does not repeat an unqualified Linux-only improvement.

AG;DR

Decision and isolated mechanism

• Documentation, evidence and tracking only. No runtime, test implementation, archive API, worker policy, CLI, toolchain or benchmark default change.
• The private candidate keeps owned ancestor keys, default mapped-node construction and subsequent assignment. Only ordered lookup/missing-ancestor insertion changes: lower_bound supplies the existing position to hinted try_emplace, with comparator equality and the stored-member strict-prefix invariant preserved.
• This is distinct from O27a deferred ancestor ownership and O27b direct mapped-value construction, already recorded by Record archive insertion qualification and reconcile tracking #1126. Mark O27c evaluated/rejected; all three requested O27 mechanisms are now independently evaluated. Do not adopt the candidate or automatically repeat it unchanged.

Results and qualification boundary

• Both hosts pass all 28 complete construction/destruction fixture metadata/path/directory oracles and three optimized archive targets. The screen is one three-repetition population per host/participant, retaining the fastest two, with a byte-identical baseline repeat. Linux uses one pinned allowed CPU; Mac is unpinned. Both use the same Clang 23.1.4 release configuration and source inputs except the intended archive implementation.
• Linux shape medians improve 2.89–4.35%. Mac results are mixed: wide-short +10.53%, flat +3.46%, deep-long -3.71%, chain -4.40%. Mac flat/100,000 costs 13.97% versus identical baseline repeat -0.79%; preserve that cost and the noisy tiny cells. The changed ancestor branch does not execute for flat members, so no causal explanation follows from source inspection alone.
• The predeclared cross-platform screen gate fails. No full three-session nine-repetition/retain-seven campaign, native/listing expansion, full-root suite or sanitizer qualification is claimed. No performance measurement in this artifact is a qualification of the current public PR context.
• The first Mac setup failed before timing because copied images lacked executable mode; the mode-only recovery and failed logs remain preserved. No successful screen was repeated until favorable.

Evidence and publication

• Add docs/benchmarks/archive-ancestor-hint-20261107.md and its 115,195-byte .json.gz, SHA-256 8359b99d815469270bbdb90aef7ec8dd11f9156ef37d6b7c5a879672badfaad7. All 53 UTF-8 source/raw/log record byte counts and SHA-256 hashes independently verify, preserving source/image identities, complete cases, copied controls and setup failure.
• Update TODO and performance analysis, register report HTML and supported exact raw download mapping. No CHANGELOG entry is needed for an unselected research prototype; the gzip is below the large-file threshold and requires no new exception. No historical performance data is removed.

Validation and dependencies

Normal signed commit b1d92ee275f2c7a1c1f5308291d4149a3771636c and normal push succeeded with all configured hooks intact. Trunk, the exact changed-file policy/format gates, canonical Markdown table checks, release-site configuration, gzip decoding, all 53 record identities and diff checks pass; published payload hashes remain unchanged. No local clang-tidy executable ran.

Based on #1128 head 081d5be6627d004f2983837d1221399559e36910, following #1127/#1126 to preserve overlapping tracking and site-publication changes. Merge the ready parent unchanged first, then synchronize and require this PR's own current main-context CI. Historical experiment logs do not substitute for its required checks. The four unique unqualified worker drafts remain untouched.

The ancestry was synchronized and pushed through normal signed merges after CI found three narrow evaluate.cc lint issues in #1127 and the required Brotli byte-pointer cast annotation in #1128. This PR's own implementation/evidence payload is unchanged. The corrected current head requires its own full CI; successful ancestor or previous-head checks are not substituted. Healthy reachable collectors were not cancelled, and local clang-tidy was not run.

Current main synchronization

#1128 merged unchanged as 1278c8943a6abd236a2f20e3937cc8e23a0f95a7. After automatic retargeting, synchronized this child with signed merge 73de2769f5a06b45c1677de7cabc31d1169d0854 and pushed immediately through normal hooks. Proved that the squash main tree, qualified parent 334cc891 and already-included parent d89fa1d7 have identical full trees. Resolved two adjacent documentation conflicts from the child only after this proof; the complete resulting tree is exactly the previous 62030d1d child tree. Runtime and all frozen evidence bytes are unchanged. The earlier stacked Test success is early feedback only; this pushed head must pass its own new main-context checks and the predecessor exact-main Test must be healthy. No local clang-tidy was run.