Obrigado — feature substancial e bem estruturada: separa qualidade operacional (comportamento de wire: 4xx/5xx, 429, respostas malformadas, stream interrompido) de qualidade semântica (só setada por avaliadores externos, nunca inferida do sucesso HTTP), com confidence/sample-awareness para não deixar poucos sucessos de sorte dominarem o ranking. Instrumentação de streaming (TTFT/ITL) threaded até RoutingEvent, endpoint de explicabilidade, e teste E2E determinístico cobrindo degradação→recuperação→blip.
Validação (worktree própria a partir de origin/release/v3.8.50, merge limpo, 0 conflitos):
- typecheck:core limpo, complexity/cognitive-complexity dentro do baseline
- 59/59 testes passando (mlx-provider, routing-adaptive-e2e, routing-events(-concurrency), routing-otel, routing-quality, routing-scoring-quality, stream-timing, auto-combo-scoring-clamp)
#7847 reports a 3.05 MiB request (729 messages / 86 tools) reaching ~12,282 MiB of V8
heap, and asks for "a regression benchmark that records peak heap for representative
500-800-message, tool-rich requests" before any fix lands. There is currently no memory
baseline in the repo at all (bench:compression is the only benchmark), so a clone-reduction
change could neither be justified nor regression-guarded.
npm run bench:heap-body attributes retained heap to each copy the chat path makes:
| mechanism | call site | retained | x wire |
| cloneLogPayload (unbounded) | chat.ts buildClientRawRequest | 3.18 MiB | 1.04x |
| cloneBoundedForLog (bounded) | requestLogger.logClientRawRequest | 0.04 MiB | 0.01x |
| structuredClone x3 (combo targets) | combo.ts attemptBody | 9.53 MiB | 3.12x |
| JSON.stringify (token estimate) | combo.ts estimateTokens | 3.06 MiB | 1.00x |
| per request (sum) | |15.81 MiB | 5.17x |
It measures the real production helpers rather than reimplementations, so a change to the
log bounds or the clone strategy is reflected directly.
Design notes:
- Deterministic: fixed-seed LCG, no Math.random(). Verified byte-identical across three
consecutive runs — without that, a before/after delta measures noise, not the change.
- Corpus lives in its own side-effect-free module so the unit test can import it without
booting SQLite (requestLogger transitively opens the DB at import time).
- Hermetic: DATA_DIR is redirected to a temp dir before importing, so the benchmark never
touches the operator's real ~/.omniroute store.
- Node, not bun: --expose-gc and V8 heap accounting are the measurement; another engine's
heap number would not describe the production runtime.
- --max-retained-mib exits non-zero, so this can become a CI gate once a target is agreed.
Reports only; wires nothing into CI and changes no production code.
