feat(loops): seeded promotion gate + authored-path convergence

[drewstone]

What

• promotionGate as a package primitive (src/runtime/promotion-gate.ts, exported from ./loops): the promotion decision over a holdout BenchmarkReport — SEEDED paired bootstrap via agent-eval heldoutSignificance, deterministic verdict, minimum-evidence floor (default 6 paired tasks), CI lower bound must clear the threshold. Replaces the bench-local unseeded pairedBootstrap gate, whose verdict was non-deterministic re-run to re-run and accepted n=2.
• authorStrategy hardening: named fallbackModel retry (one attempt when the primary call fails or returns no code block — the edge-524/thinking-model case), and temperature/maxTokens are now actually passed through (temperature was declared but silently ignored).
• assertAuthoredCodeSafe → assertStrategyContract: same checks, honest framing — the lint enforces the harness's two measurement invariants at the module boundary (author blindness: no out-of-band reads/mutations of verifier state; conserved dose: no out-of-band compute), which is what keeps harness-verified scores and equal-budget comparisons meaningful. It is a contract lint for measurement integrity, not a security boundary.
• Bench convergence: bench/src/strategy-author.mts drops its duplicate authorStrategy + drifted contractDoc (it lacked the fix(runtime): authored-code import enforcement + empty-messages foot-gun #219 empty-messages rule) and becomes the R0→R2 ladder CLI over the package primitive; flywheel-run.mts authors and gates through the package. Authored run artifacts are gitignored + excluded from bench typecheck.
• Regression tests (tests/loops/strategy-suite.test.ts): harness-verified scoring (a {score:1}-doing-nothing body scores 0; keep-best overrides under-reporting), the messages: []-is-fresh rule, the contract lint's accept/reject set, and the gate's determinism + floor + fail-loud pairing.
• bench/HARNESS.md synced (author default deepseek-v4-pro, rotating holdout, packaged gate).

Why

The flywheel is the measurement instrument for the self-improvement program; its promotion verdict must be reproducible and its author path must have one implementation. This closes the bench/package duplication (drifted contracts), the unseeded-gate hole, and puts the #217/#219 correctness claims under test.

Verification

• pnpm run typecheck clean, pnpm run lint clean (13 pre-existing warnings untouched), pnpm test 698 passed / 1 skipped (+18 new).
• cd bench && npx tsc --noEmit -p tsconfig.json clean against the rebuilt dist.

[tangletools]

✅ Auto-approved PR — 5151694e

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-10T10:17:27Z}