FLARE-FIX-001 incremental review package

flare-fix-001-incremental.patch is the reviewable delta from the saved local baseline to the reviewed v2 source. It has no private paths and is not an upstream submission or standalone cleanup patch.

Exact base and replay

Start at 47581ccac82660953a5b59cb04b2750d64edc97d, apply ../baseline-source/pre-existing.patch, and install ../baseline-source/scheduler_trace.rs as helios/src/server/scheduler_trace.rs. Then apply flare-fix-001-incremental.patch. A fresh temporary replay applied this sequence and matched every reviewed-current file byte-for-byte.

Scope

The delta adds the configurable row/page allocation split, defaulting to 25%, and the opt-in transfer-byte audit with its configuration, budget accounting, tests, and docs. It changes 12 repository files, with 1,022 additions and 257 deletions.

Prerequisite dependency

The patch depends on the saved pre-existing local telemetry/instrumentation baseline; it is deliberately excluded from this delta. In particular, the existing StoreStats host-memory fields and their producers in helios/src/offload/{composite,index,memory,mod}.rs are consumed by the existing metrics/runtime wiring, and the unchanged scheduler-trace module and its server wiring remain present. This package does not claim those prior local changes are standalone cleanup or ready for contribution.

For future review, keep the prerequisite telemetry/scheduler-trace work as one separate unit. The retained v1/v2 snapshots now support the exact, review-only sequence in SPLIT-REVIEW.md. The original combined incremental patch remains the authoritative replay artifact; the two split patches are not newly tested or proposed for submission.

Review boundary

The byte audit proves transfer payload equality for its exercised scope. It does not resolve the failed sampled-output gate or provide the unimplemented teacher-forced/logit numerical oracle.

The completed three-pair experiment supports a workload-specific cache benefit: the complete 52-request one-token reverse-replay populations match, and their median first-generated-token latency falls by 95.038–96.065%. Row-pressure events fall from 112 to zero over priming plus replay, while completed eligible restores rise from 10 to 36 per measured replay. These are allocator diagnostics, not an engine comparison or a qualified default recommendation. The numerical gate still fails, and the third pair exceeds the protected regression limit for the small working set and page-heavy lane. See the hash-linked causal assessment and independent review.

Keep the existing 25% default. The startup option is the useful contribution shape to discuss with maintainers; the tested 75% setting is an explicit workload choice requiring further numerical and regression work. Show this package and its limitations to the user before any upstream submission.

At the 2026-10-23 UTC freshness check, upstream main was e5bd6a67, 132 commits after the experiment base. The named upstream surfaces still have the quarter-row allocation and no equivalent row-share/admission options. Adjacent PrefixStore::poll failure handling and Orbit forwarding changed; an eventual contribution must reconcile those contracts against then-current upstream. The experiment remains pinned and has not been rebased. See the bounded official-source comparison.

Later test-only delta

After the v2 image and patches were frozen, ../lifetime-test-proposal/lifetime-test.patch added one composed CPU Worker/scheduler cancellation test. Apply it after the v2 combined delta if reviewing this later coverage. Its SHA-256 is 43f14e55bd3724c87d9a04c847ee50d2412ed2df056b5826f52118faa4fde149; focused execution with server,runtime-package passed in ../lifetime-test-execution/r2/execution.json and was independently accepted in ../review/lifetime-fixture-review.md. The original missing-feature compile failure remains retained. This test-only delta is not part of the frozen v2 Docker image, original combined patch or split-patch replay claim. It covers controlled CPU success/failure completion after shared-waiter cancellation, not real DMA overlap or model numerical fidelity.