Finish full-suite verification for BuildTest resource estimate reuse

Written 2026-10-27 04:08 UTC.

RE-VERIFY: All state below is a write-time snapshot. Recheck GitHub branch heads, draft PR state, and result XML before continuing. CI has not been watched in this lane.

Required behavior

The owner asks to "keep a versioned shared resource-stat projection updated incrementally for the affected identities/projects when new telemetry commits, while dispatch still sees fresh data."

The owner requires: "Never merge, enqueue, deploy or publish." Never raise a timeout or limit, reduce iterations, disable tests, or edit production state. The brief permits one complete-suite race, and this lane has already attempted it.

Mission

Obtain complete-suite verification for the implemented dispatch projection and finish review of the draft PR. A new complete attempt needs orchestrator direction because the one permitted race ended without usable results.

What was found and done

The owner reported unreliable visits to droplets.wasmserver.com. The CPU investigation measured BuildTest 31.2%, ContainerNursery 11.6%, and kotlin-build-ci 7.7% of host CPU. BuildTest instruction samples included 221 hits in currentPeakMemoryEstimatesBySpec and 55 in currentCpuDemandEstimatesBySpec. Owning findings.

BuildTest globally normalized/grouped memory and CPU history and recalculated project medians for every relevant runner event. The new public API test reports five calculations for unchanged generation zero, where one is expected, on all three baseline attempts. The current expanded scenario gives the same result on all three reversal attempts.

The fix shares one versioned dispatch projection. Commits mark affected identities; a read copies those identities under duration ownership, calculates outside that ownership, and retains later concurrent commits for the next read. Only changed memory values update their project medians. Per-run memory and CPU maps reuse the projection version. Rotation, live expiry, backward time and restart are covered. The independent public statistics cache keeps its established contract.

The scenario uses the real embedded service, the published in-process SSH component, real runner queues and archive streams. It passed without retry in 27157 ms, again after the latest-main rebase in 22204 ms, and after the immediate pre-PR rebase in 22718 ms. It covers exact identity/project work counts, fresh memory and CPU reservations, project medians, restart, live expiry, backward time, immutable observations, a nested statistics read and observation-error recovery. Independent test and correctness reviews completed. No existing test was edited.

The full local/remote race used the exact source commit below. Remote provisioning ended with an unregistered droplet-service lookup; its 5349 time-zero rows were synthetic. Local execution ended with code 137 and no XML; its initiating cause remains unknown. Complete-suite verification is NOT met. See the suite verification gate for the remaining condition.

Relevant refs

Deployed or published state

Nothing from this lane was deployed or published. No deployment, restart, route update or Maven upload command was invoked. All feature and control work remains unmerged on the linked branches. Both source checkouts have no tracked edits or stashes, and remote branch heads match their local commits. The unused initial transfer directory was removed. Generated jars and caches are excluded from the evidence.

Next steps

1. Read the suite-gate dependency and ask for orchestrator direction before another complete run.
2. Preserve the first-attempt requirement and existing timeouts, limits and iterations.
3. Use the existing public test command: scripts/test.bash --local --test tests/testDispatchReusesUnchangedResourceStats.kts --log targeted.xml.
4. When broader validation is available, review the draft PR; do not merge or deploy under this brief.

Operational notes

Full findings and captured verification records hold the exact commands and result counts. The expanded scenario is byte-identical between feature and reversal checkouts. Clone main before fetching a WIP branch, or fetch sufficient ancestry before rebasing a shallow branch clone; the initial control rebase reported an add/add conflict until ancestry was present. Temporary setup errors are distinguished from test attempts in the findings.

Review note: the old rotation clear/reload order predates this change. An unsuccessful reload records an unfinished publication before either resource generation advances; the new view still describes the prior accepted generation. The publisher schedules that unfinished transaction again. This concern did not identify a newly accepted telemetry generation missing from the new projection.

Open question for the orchestrator: may one additional complete-suite race run once the execution conditions are understood and resolved? The original one-race allowance has been used.

Final PR creation: https://github.com/CodexCoder21Organization/BuildTestEmbedded/pull/1681. It was created as a draft. No CI status was read after creation.