epic: development speed -- merge queue, generated derived data, prebuilt binaries, affected-only runs, delivery metrics #6595

Development is slowed less by writing code than by waiting and redoing work around it. This epic tracks the fixes, ranked by measured cost. Evidence comes from the 2026-10-05 overnight run (ledger rows on #6063).

Observed bottlenecks

1. Master is red most of the time (spec-guards, coverage, duplicate-bodies, lean, fpga-conformance), so every PR spends effort proving "same failure as master".
2. Derived data (seals, the t27b ledger) is checked in and goes stale. Two standalone bless PRs (ledger(t27b): bless t27b_expectations.json from lab run 4040407cd (Refs #6063) #6565, ledger(t27b): bless from lab run 3313df937 (Refs #6063) #6576) went stale within an hour, and 30 stale seals turned every bee PR red.
3. The t27b lab measures only master, under qemu, so PR branches cannot be measured before merge (t27b: native arm64 CI per PR on ubuntu-24.04-arm and macos-14 (off the Mac, off qemu) #6444).
4. t27c is rebuilt from scratch in every task by every agent.
5. Lanes serialize on hot files (cli/t27b/src/lower.rs).
6. About 30% of bee runs end with no diff, and the cause is not logged.

Plan (practice -> where it comes from -> issue)

• A. Merge queue + required checks: test the merge result, not the branch (Rust bors / GitHub merge queue, Chromium CQ). OWNER: repo settings.
• B. Derived data generated by CI, not hand-committed (Bazel/Chromium style). Seals and the ledger are computed, and PRs stop editing them.
• C. Prebuilt t27c/t27b binaries per master commit as artifacts, plus a build cache (Zig, Go, Rust nightlies).
• D. Per-PR native arm64 runs (Go trybots, LLVM pre-merge; GitHub ubuntu-24.04-arm): t27b: native arm64 CI per PR on ubuntu-24.04-arm and macos-14 (off the Mac, off qemu) #6444.
• E. Run only affected specs, keyed by the use closure that the reference cache already keys on (Bazel/Chromium test selection).
• F. Automatic reducer and bisection (Go bisect, llvm-reduce): t27b: automatic reducer for blocked and disagreeing specs (tri t27b reduce) #6445.
• G. Split hot files so lanes run in parallel (lower.rs already partly split).
• H. Delivery metrics in tri: lead time to merge, change-fail rate, time red, and why bee runs end empty (DORA).
• I. Self-host t27core so compiler work stops needing foreign-code exceptions: epic: self-host t27c -- rewrite the compiler from Rust into t27 #5980.

The only-t27 rule applies: decisions live in .t27, and generated code is allowed; hand-written glue needs the owner label.