Skip to content

Todo summary-display follow-up qualification (2026-10-27)

This entry supplements the original retirement record. The production boundary remains summary display through todo.context.page and in-process quota planning; both complete planning decisions remain necessary. There is no new provider, format, permission or RPC owner.

Preserve the earlier evidence

The independent review of head c242ca5f439c269c313329c883299a28d95a55dc against be1dc37a008be7017b64b622f653b6707df5c544 found:

  • the registered todo_decision_scope_outcome_v0 correction passed canonical File/SQLite source and installed wheel/sdist checks;
  • one SQLite CLI fixture used 83 → 82 requests, including 22 → 21 todo.context.page calls. Its workload differs from the original N1/N3 75 → 74 trace; neither trace permits removing another planning decision;
  • a fixed 32-pair CLI cohort had base/head p50/p95 336.97/521.98 and 345.96/836.79 ms, a 60.31% candidate p95 increase;
  • the predeclared 64-pair rerun had 301.56/400.98 and 305.88/391.08 ms. This later result does not erase the earlier failed cohort. The difference did not establish a repeatable regression or explain the earlier tail;
  • cold startup, warm typed requests and idle/burst daemon RSS were still missing. Python thread profiling did not measure Node CPU or establish tail causality.

These historical cohorts keep their original status. They are not relabelled as passes, and the default SQLite limits remain unchanged.

Rebase and real recovery

The follow-up baseline is ff8d9932294e9004b9ffeb2fb473d00354397eb8. Six File/SQLite recovery test cases initially stopped on canonical_authority_required, before their intended assertions. The fixture now initializes the real provider from authored source facts before calling the production CLI; it does not restore a legacy read fallback. All eight receipt/next-action cases pass, including version preservation, rejection of wrong Todo step updates, legitimate original-Turn settlement and replay. The earlier six failures remain recorded as fixture failures, not proof of previously passing recovery behavior.

Final-head measurement contract

PR #7411 records the immutable final head, execution commands, fixture fingerprint, environment and results. Its bounded offline qualification uses the same synthetic canonical SQLite Todo and interpreter on both source revisions, with no active Goal mutation:

  • 24 managed-process cold-start pairs, distinguishing warm filesystem/compile cache from a fresh process;
  • 64 complete quota should-run CLI pairs with alternating AB/BA order, all samples retained and nearest-rank p50/p95;
  • 256 warm typed requests and 256 complete quota-planning projections per revision, separating transport from the Python facade;
  • idle, sampled burst and post-burst daemon RSS over 2,000 bounded projections;
  • real CLI identity/result assertions, focused TS/Python coverage and packaged readback, reported separately from timing.

Do not overlap owned heavy builds/tests with timed cohorts. Keep private raw captures outside Git and publish only synthetic aggregates. A single synthetic workload does not establish sustained large-Goal performance. Any miss retains an owner-review hold; a faster microbenchmark cannot waive a full-CLI failure.

Rebase onto the subsequently merged main

The final integration base advances to 17b2a0f9b818dc6e6fe77ccc567b20ff91b3bc40. Main now includes #7459's canonical fixture repair and changes to the real quota/context path. The conflict resolution keeps that shared fixture repair instead of duplicating it; this PR retains the File/SQLite receipt-version regression. The earlier ff8d9932294e9004b9ffeb2fb473d00354397eb8 comparison and e7a54b8f32a648124133b02e11404b75a421fdac head remain historical evidence. The final PR reply records fresh exact-head correctness, package and matched performance results for the new base. No earlier cohort is rewritten.