perf: measure large trackers and reduce native mutation overhead
Held ownership contenders previously decoded the complete corpus before learning they should skip or refuse. Claim/release now prepare lineage under the item lock after journal recovery and atomic eligibility checks, preserving the refusal contract while avoiding repeated corpus scans for rejected claim-next candidates.
| Finding | Result |
|---|---|
| Contended claim-next scans | Real locked competing mutations after ranking cover 10/100 attempts, if-available skips, hard refusals and acceptance on attempt 100. Exhaustion reads the corpus once; acceptance reads it twice total. |
| Refusal precedence | SDK and actual CLI regressions retain held exit-0 skips, already_claimed_by exit 4, and missing-ID empty-author/future-clock argument errors at exit 2. |
| Lock/recovery ordering | Busy locks precede unrelated corruption; lost-item journal replay occurs before lineage validation. The snapshot is taken under the target lock. |
| Oracle pin coupling | Installed-version validation follows the benchmark manifest, after executable fingerprint refusal. Independent oracle metadata pins are covered. |
| PRRT_kwDOTwgMRs6vX_PV | Profile accepts only yes/no before binary reads or fixture creation; true/maybe/false are rejected by a real-process regression. |
| Stale measurements and coverage | Complete current source receipts, sample tables, coverage totals, source digests and contention observations are committed. Older tables are explicitly historical. |
| Future source-PR workflow support | Remains the open follow-up pm-rust-bench-workflow; the branch gate, fixed historical baseline and implementation-specific proof are scoped to this PR. |
Measured implementation: 57845abdf2beb9087bfca25e674d56c32805a608. Final receipt head: 069a325f3485261d29bd5c46376c90bfecd4a721. Its diff changes only docs and PM records; all 113 tracked implementation/harness inputs have identical digests.
Final receipt head passes CI 39864879374: 9/9 jobs. Each unchanged just release-check passes ordinary and instrumented suites with 100% lines, functions, regions and branches: Linux 589 + 589, macOS 578 + 578, Windows 572 + 572, all with zero failed or ignored tests. Final-head benchmark/proof 39864879217 passes 2/2 jobs. The committed measurement implementation also passed its 9-job gate and complete paired benchmark/proof.
Seven behavioral negative controls fail the intended assertions (four Rust exits 101, three Node exits 1), and all seven restored controls pass. The two original optimization controls also fail when only their behaviors are reverted and pass after exact source restoration. Commands, test names, source hashes and outcomes are retained in the committed evidence. The complete paired matrix has 81 rows and 1620 measured samples with distinct before/native executables; it preserves all 27 baseline/native outcomes and all 9 write artifacts. The rebuilt final receipt-head native and baseline executables have exactly the same SHA-256 fingerprints as the committed matrix.
The 50000-item contention diagnostic uses real non-git trackers and actual competing ownership writes. Its single debug-process observations include competing writes and ranking; they are not fresh-process percentiles. The scheduling hook is absent from production builds. Existing native/oracle gaps remain unchanged: native succeeds 19/27, output equality is 2/27 and write-artifact equality is 0/9. The allowed RUSTSEC-2024-0449 advisory and two existing tracker health advisories remain.
Local validation passes 9 ownership contracts, the five-variant contention regression, 25 Node harness tests, 7 committed-head identity audit tests, strict erasable TypeScript, changelog generation/check and strict tracker health. The local full gate ended at the existing five-second relationship_large_edge_limit_finishes_at_budget_floor deadline (485 passed, 1 deadline failure); its targeted retry hit the same deadline. The exact unchanged test passes all CI ordinary and instrumented suites. Two earlier Node nested-process timeouts passed on the complete 25-test retry. One PM startup overlapped oracle installation and was rerun successfully after setup; no local full-gate pass is claimed. Thresholds and source scope are unchanged.
Fresh retrieval finds only the unresolved profile thread PRRT_kwDOTwgMRs6vX_PV; its fix and regression are committed. The latest substantive CodeRabbit review is 2026-10-22T20:53:52Z at 9337cb1 and contains no outside-diff findings. The current CodeRabbit check requires manual review, so fresh approval remains unavailable. No thread was replied to or resolved by this work. No skipped or manual reviewer check is treated as fresh approval.
PM: pm-rust-wdj4, pm-rust-bench-workflow. The implementation item remains in_progress, its claim is released, the follow-up remains open, and no item is closed. All 76 pre-existing history prefixes remain intact.