Repository navigation
Avoid decoding the same captured streams twice during native trace construction #4537
Description
Native observation trace construction decodes each captured stream for initialization and then decodes the same stream again for indexing. Both operations consume the same immutable captured prefix within one construction. Repeated decoding adds avoidable work as history grows; successful scanner microbenchmarks alone have not resolved the native acceptance deadline.
Resolve by reusing decoded rows within a single trace construction while retaining initialization, model/session, ordering, callback, decision and result validation. Public initialization behavior and error ordering must remain compatible. Do not add a persistent cache or reuse data across distinct captures.
Validation should cover malformed and contradictory initialization, changed captures, independent trace instances, and a direct count proving each stream is decoded once per trace. Final native acceptance remains required; this issue does not claim the duplicated parsing alone caused the deadline failure.
Activity
realmarcin commentedon Oct 6, 2026
The duplicate decoding is reproduced by a direct constructor count: the original path decodes both streams twice. The local correction reuses decoded rows within one trace construction for initialization and indexing, retaining all validation and avoiding shared state across separate constructions.
77distinct regression checks are satisfied, including malformed initialization, changed captures, independent trace instances and transcript-before-control error ordering. Independent source review passed. Complete-constructor timing and final native acceptance remain pending; this is not yet a released fix.
realmarcin commentedon Oct 8, 2026
The reconstructed single-decode candidate is published in #4659 at
a8c11973da120cc0f792eef2d574dab7d910b09d, based on the available0125adapter. The independently reviewed comparison is published in #4664. Complete canonical constructor state, method results and refusal graphs match on the identical frozen capture; decode counts fall from two to one per stream. All 90 focused candidate tests and 184 comparison-driver tests passed. Median constructor thread CPU across three alternating pairs was 101.3 ms versus 54.8 ms, with all six samples and first-sample/cache limitations retained. This is not a whole-path or native-acceptance result.Required CI prevents integration. In run 37745542983, the neutral full offline acceptance test reached the unchanged 900-second registered deadline. The omission test separately refused an aliased authority. #4665 tracks a proven shared-schema hardlink mutation in an unrelated concurrent regression test, whose timing supports contamination as the omission failure explanation; the exact offending path was not retained, so that attribution remains an inference. The fix will isolate both hardlink endpoints inside the fixture and retain the production rejection rule.
The two failures are retained separately. Correcting test isolation does not resolve the deadline or justify relaxing any acceptance gate. #4537, #4354 and #4400 remain open; no real native/provider campaign has run.