[helly25]

Reconcile an archive B-tree experiment that had already finished but was still listed as unmeasured. Preserve the historical observations and reject that container substitution because construction gains came with slower wide/deep directory queries. Runtime behavior is unchanged.

AG;DR

Scope and dependency

• Report-only child of Record rejected queued read cancellation policies #1132; merge the existing Reuse validated rg output transactions in matcher results #1127 -> Decode embedded language JSON into final owned storage #1128 -> Record rejected hinted archive ancestor insertion screen #1129 -> Stop main benchmarks once three of four candidates agree #1130 -> Record deferred immutable ZIP locator evaluation #1131 -> Record rejected queued read cancellation policies #1132 chain first. Shared tracking, performance history and release-site configuration require sequential integration. No existing parent head is changed or delayed by this child.
• Exactly six files: historical report and lossless capsule, TODO, performance analysis, release-site page/download mappings, and a filename-specific large-file exception. No runtime/header/test/BUILD, dependency, new private draft, worker or compiler change is published.
• Correct the stale O19k unchecked/unmeasured entry and future-audit prose. Completed status means evaluation and rejection of this exact historical global-container policy, not adoption. A new directory-specific index remains separate work.

Historical decision and protocol

• The preserved candidate replaced only the private archive node std::map with absl::btree_map, using the pinned default transparent three-way string comparator and direct dependency. Index-only mutation and nonescaping read iterators were audited; returned metadata/text/member sources own their data. No borrowed-key or insertion-hint policy is folded into this trial.
• Each host has three balanced sessions (parent/candidate/parent, candidate/parent/candidate, parent/candidate/parent), nine raw repetitions per invocation, fastest seven averaged. Duplicate roles combine geometrically within a session; per-case results are medians of the three candidate/parent session ratios. Linux pins CPU4; Mac does not claim affinity enforcement.
• Complete corrected populations contain 28 construction/destruction cases and 88 directory/metadata cases per host. All eighteen unique JSONs and eighteen logs per host survive. Initial runs that overwrote one same-role control remain explicitly incomplete exploratory data and are not pooled with the corrected population.
• Construction medians improve 7.06% Mac/0.66% Linux, while directory medians regress 1.92%/2.60%. Wide medians regress 3.71%/6.20%, deep 2.35%/7.87%; all slower cases and per-session ratios remain visible. Strong flat/chain wins do not erase these costs. The global container substitution stays rejected.

Provenance and qualification limits

• Historical source parent is 45f4c5e65bfb7de5d0facf8bcbdca1d28320ae4a. Current experiment-planning base d8cf61cccd4b0efed5058bf6fa45228e47162b48 has different archive ownership/scanning and release/link settings. Old percentages are not attributed to current production or added to later gains.
• The original decision note reports optimized archive filesystem tests passing on both hosts and about 39 KiB benchmark-image growth. Complete contemporaneous compiler-action/source/image receipts were not frozen; later captures are labeled later and cannot prove original measured executable bytes. No full target-suite, sanitizer, product-size, RSS or current-main runtime qualification claim.
• A duplicate private draft was stopped before compilation/test/timing after the old evidence was recovered. Its drafted source/test stays private. This PR runs no new performance trial.

Evidence and publication checks

• docs/benchmarks/archive-btree-historical-decision-20261107.md links the adjacent capsule through supported release-site files mappings.
• Capsule: 2,205,832 compressed bytes, 155 individually hashed records, 43,505,320 decoded bytes. SHA-256 e609682737ebd09c27b5760e3dd811e55f7e54fbc9c1b852c7e21f8dcf983df0. Includes corrected/initial populations, original note, source patch, analysis/drivers and expressly later identities; executable images stay private.
• Independently verified every record length/hash and all four handoff files; report/capsule remain byte-identical. Recomputed 232 per-case session medians and the four construction/directory cohort medians, matching the report.
• 32 release-site tests passed; configuration/page/file and exact large-file allowlist positive/negative controls, Markdown tables and diff checks passed. Trunk formatting and normal signed commit hooks passed; local clang-tidy was not run as instructed.
• Normal push hooks passed. This PR's own current-head CI remains required. Early stacked checks cannot replace post-retarget main-context validation.

Current main integration

Signed synchronization 758c69f067915c515d60d0496047218adc0cae26 incorporates #1132 main squash 5d413a35cbecbe81d23f6b4bfd95b3d95840e484. The qualified child tree remains exactly 492e750ab77c0abd7fb0e62f4ad5c207d45b3e30 after proving the real included old parent, the qualified signed parent and actual main have equal full trees. Normal commit and immediate push hooks passed; merge hooks intentionally skipped redundant Trunk. No local clang-tidy ran. This pushed head requires its own fresh current-main CI; earlier stacked success does not establish current readiness. Exact predecessor-main Test must also succeed before merge.

Final current-head qualification

Own current-main Test run 42378590920 succeeded on signed head 758c69f067915c515d60d0496047218adc0cae26, including all ordinary tests, hosted site validation, CI clang-tidy, coverage, sanitizers, both-platform measurement agreement, best-seven comparison and done. A fourth Mac candidate completed successfully; after-fourth agreement passed and fifth candidates skipped normally. Exact predecessor-main Test 42378502033 also succeeded. This pushed head is ready and is merged unchanged under the user’s requested #1133 → #1134 order. Historical research measurements remain separately scoped; no current-runtime performance claim is added.