ci: gate concurrent retention and logging throughput

[codeforester]

Summary

The comparative benchmark now gates concurrent retention and sustained logging. Twelve synchronized processes share a warmed cache, and persistent/ephemeral logging emits 3,000 records per sample. Reports use schema v2 with explicit latency, throughput, ratio, and budget units; the retained CI artifact remains compatible by filename.

Issue

Fixes #391

Branch maintenance

PR #421, the Project Intake parent car, has merged into main. GitHub retargeted this benchmark PR from the parent branch to main; the benchmark changes remain unchanged in the current head. This car is now the first open implementation car in the 0.5.0 train, followed by the release candidate in #423.

Current-head validation

At 0418cd45a59a9465934e85a6ed1edbb1f7759ffa: the branch includes the current main and the benchmark gate changes. The prior local validation covered uv lock freshness and baseline, runtime, strict typing, style, and contracts with all declared extras: 660 passed, 1 warning, and 294 subtests passed in 10.63s. Hosted checks are being rerun after the retarget; report their live result from the Checks tab before merge.

[codeforester]

Follow-up on current head 1cddd44: the new benchmark gate was calibrated against retained hosted artifacts. Functional/platform checks were green; the Unix persistent-logging p95 was 55.58 us/record and the macOS repeats reached 129.66–160.78 us/record due hosted filesystem variance. The documented caps are now Unix 80 and macOS 200 us/record, with the original ~107 us/record regression signal still below the caps. Benchmark unit tests and Ruff pass locally.

[codeforester]

Addressed the benchmark-gate concern in the pushed head 258f90b0.

The stress run now measures plain stdlib FileHandler logging in the same process and requires a persistent-to-stdlib p95 ratio in addition to the existing absolute cap. The macOS absolute cap remains a filesystem sanity ceiling, while the ratio is the regression signal. The ratio cap is 12x across profiles; the full 31-sample local macOS run measured 63.44 us/record persistent, 6.36 us/record stdlib, and a 9.98x ratio.

Validation: the supported benchmark gate passes with 31 samples, and the contracts/style gates pass.

[codeforester]

Resolved the merge conflicts by merging current main, including merged PR #421, into this branch in commit 0418cd4. Preserved the #391 benchmark report v2 and contention/logging gate documentation, the Project Intake recovery section, current main logging sidecar recovery, and the combined hot-path tests. Local validation passed: benchmark/CI-focused tests, full pytest, Ruff check and format, strict mypy, and documentation validation. GitHub reports the PR as conflict-free; hosted checks are running on the new head. The PR remains open and unmerged.

[codeforester]

Investigated the failed Benchmark (unix/windows) checks. The failures were filesystem-tail outliers in the pre-existing persistence scenario: Unix reached 588 ms p95 with a 15.7 ms median, and Windows reached 1123 ms p95 with a 54.6 ms median; all #391 concurrency and logging-throughput gates passed. A Windows rerun passed, confirming runner variance. Added commit b486c93: Unix retains the 50 ms sustained median guard and now uses a 750 ms p95 filesystem-tail sanity ceiling, backed by the retained artifact and regression tests. Local full pytest, benchmark tests, Ruff, strict mypy, formatting, and docs validation pass. Hosted checks are running on the new head; the PR remains open and unmerged.

[codeforester]

Hosted confirmation: the fresh benchmark run for b486c93 passed Benchmark (unix) and Benchmark (windows); macOS remains queued. The Unix persistence-tail calibration no longer blocks the gate, while the sustained median and #391 logging/concurrency ratio checks remain enforced.