feat(submitqueue): code-aware jev scorer, scoring benchmarks, and a speculation liveness fix

[tvandalovsky-cw]

Why?

SubmitQueue's speculation ranks builds by each batch's probability of succeeding, and the shipped scorer prices every batch at a flat 0.5. This branch explores a code-aware scorer backed by a System One decision model (TypeSafe Jev, reached through a LiteLLM proxy), and builds the infrastructure to measure whether it beats simple baselines: an oracle ceiling, a real-history replay benchmark, and a synthetic corpus with build-verified labels. Along the way it fixes a speculation liveness bug the experiments exposed.

What?

Three commits, reviewable independently:

1. fix(submitqueue): cancel undispatched paths in the run that breaks them — a funded path whose build was never dispatched stayed cancelling after its guess broke, holding a budget slot until a later speculate run that a quiet queue never sends. At buildBudget: 2 this wedged the queue with batches stuck in speculating. Such paths are now cancelled in the run that marks them. Includes a regression test that fails without the fix. Self-contained and suitable for upstream on its own.
2. feat(submitqueue): add jev scorer with code-aware context and replay benchmark
   • scorer/jev — scores a batch through the proxy with a shared price cache, a fallback scorer, and configurable byte caps.
   • extension/changediff — new extension returning each change's diff and, optionally, modified files in full and package-scoped references (untouched callers of changed declarations, definitions of called functions), read from the orchestrator's existing git copy. Jev sends them as an annotated repository_context.
   • scorer/oracle — benchmarks-only scorer pricing batches from recorded outcomes; the ceiling for comparisons.
   • buildrunner/replay + tool/replaybench — benchmarks-only replay of GitHub merge-queue history: fetch attempts, rebuild each PR's change on an epoch base, answer builds with the recorded CI outcome and duration.
   • RFCs: doc/rfc/submitqueue/code-aware-scoring.md, doc/rfc/submitqueue/replay-benchmark.md.
3. feat(tool): add jevcorpus synthetic scorer corpus — generates Go changes by mutating a seed repository, labels them by building and testing, writes a replay-format dataset, and grades jev offline. RFC: doc/rfc/submitqueue/synthetic-corpus.md.

Everything new is opt-in configuration; existing queues behave as before. oracle and replay log a warning when configured and are documented as never for production.

Findings so far (details in the RFCs)

• With file stats only, Jev is no better than chance; with the diff it catches failures visible in the changed lines, and with files + references it was the best scorer on a synthetic subtle-failure workload (AUC 0.79; fewest builds at budget 4 and fastest to land at both budgets).
• It still misses a cross-file signature break even when shown the broken caller — better handled by a deterministic static check outside the model.
• On a first small slice of real Jaeger history (20 PRs, 3 failures, 2 likely flaky) it did not predict failures; more data is needed before drawing conclusions.
• The current proxy deployment rejects requests above roughly 16–20KB (HTTP 502), so deployments must set jev.diff.maxDiffBytes / maxContextBytes; the code defaults are far larger.

Test Plan

• make lint, make check-gazelle, make check-tidy, make check-mocks pass.
• make test: 154/155 pass. The one failure, //tool/compose:compose_test, fails identically on main without these commits (a macOS /var vs /private/var path assertion introduced upstream in chore(docker): supply docker compose via Bazel uber/submitqueue#830), so it is unrelated.
• make e2e-git-test passes (needs --spawn_strategy=local on macOS, where Bazel's sandbox blocks Docker buildx).
• New unit tests: git differ (diffs, stacks, stale target, files, scoped callers/definitions, renames), jev request shape and budgets, oracle, replay runner and truth, replaybench verdict logic, profile wiring for jev/oracle/replay, and the speculation regression test.
• Local queue experiments: 8 scorer configurations × budgets 4 and 2 on a synthetic workload, and an end-to-end replay of real Jaeger PRs through the git provider.

Issue

None — exploratory work on this fork.