3.2 KiB
Acceptance Efficiency Repair v19
Measured Steps
The fixed 20-cycle baseline is in efficiency-next-baseline.json and .md.
It includes older provider failures; those are not evidence of a new transport
regression after v18. The largest observed phases were:
| Step | Total seconds | Interpretation |
|---|---|---|
| After source-read events | 534.7 | Observed model/service interval, not disk read time |
| First response wait | 329.5 | Includes historical failed requests |
| Editing | 204.6 | Model editing interval |
| Provider backoff | 196.7 | Includes historical retry sleep |
| Other decisions | 190.9 | Between-event interval |
| Failure recovery | 174.1 | Between-event interval after errors |
| Builds, all stages | 117.5 | Pre-turn, post-turn and direct builds |
| Actual source-read commands | 1.5 | Recorded command execution |
The 110.14-second Vite command in this sample measures a test server's lifetime, not a 110-second startup. It is not justification for killing the test server earlier or treating that interval as CPU work.
Confirmed Waste
A Part acceptance batch selected 11 business IDs, but the existing contract
covered only Part:part.createRecompute and Part:part.importExport.
The recorded failed attempt spent 23.295 seconds building, 1.128 seconds syncing
and 1.270 seconds running cases before reporting all 11 as unimplemented.
Missing-case acceptance turns could also fall through to desktop native marker instructions on resume. Their case and contract excerpts competed with native sources for the packet budget, prompting additional source reads.
Implemented Fix
- Check the actual acceptance contract, script and selected-ID coverage before building or running cases. Return an explicit missing-ID diagnosis.
- Keep build caches, pending builds, receipts and unfinished business records.
- Put acceptance case files first in the support excerpt allocation and include selected, covered and missing IDs in the packet.
- Route missing-case acceptance turns, including resumes, to case-authoring recovery with prior context retained. Native changes remain appropriate only when an actual implementation gap is found.
- Never infer business completion from a coveredIds entry or native marker. Covered batches still use the real build/sync/acceptance pipeline.
Verification
An isolated replay used the real 11 IDs from
business-evidence/1a7a0ee6a2d3e5addb505477-20260908T104625913994.json
and the current Part contract. The in-process preflight reported all 11 missing in 0.001681 seconds,
with zero build or acceptance subprocesses. Both the contract and existing
case script were retained in the generated editable packet.
Replay data: acceptance-preflight-replay.json. The live batch and build cache
were not modified by the replay. The historical build timing and the isolated
preflight are different execution states, not a whole-project speedup ratio.
Regression coverage verifies the early stop, cache preservation, routing to business rather than compiler recovery, and fresh/resumed case context.
The combined regression run passed 131 tests. Scheduler PID 1814354 loaded v19 with a matching source fingerprint and continued through cycle 5017 using the existing HTTP/1.1 model transport.