fix(tuner): wire the research-scout benchmark feeder into model_routing

[Nibbler1250]

Problem

ModelRoutingSubject exposes an injectable benchmarkProvider (the "research
scout" seam) whose default is async () => []. At composition in
registerWisecronSubjects, only the dispatchReader was wired — the
benchmarkProvider was left at the empty default.

Consequence: the benchmark reroute has no external evidence and can never
surface a new-model reroute. It is the benchmark-side twin of the obs=0
dispatchReader gap that was already closed on the cost side. The scout itself
(getModelBenchmarks — cache-first over the Artificial Analysis free tier — plus
enrichWithAnthropicCoding) was already built and tested; it was simply never
injected, so the whole proactive routing face sat inert.

Fix

• Add makeBenchmarkProvider() next to makeModeDispatchReader in
  observation-readers.ts (the runtime seam): cache-first fetch, ensures the
  cache dir exists (the cache writer uses writeFileSync without mkdir),
  enriches Claude coding-index gaps from the Anthropic seed, and is graceful —
  a missing key or benchmark outage yields [], never a throw, so the proactive
  loop is never stalled. fetchImpl/nowMs stay injectable for tests.
• Inject it at ModelRoutingSubject construction. Cache path and TTL are
  config-overridable via benchmark_cache_path / benchmark_ttl_ms.

Why it's safe

• Governance preserved. The subject still never fetches the web itself — the
  runtime composition injects the feeder, exactly as the subject's own comment
  prescribes. The subject remains pure and unit-testable with an injected
  provider.
• Graceful degradation. No API key or a fetch failure returns [] (same
  contract as the cost-signal reader); a stale cache is preferred over nothing.
• No fabricated data. The Anthropic SWE-bench seed is left untouched — I did
  not hand-add Opus 5's figure (the seed file explicitly forbids guessing).
  Once wired, the live scout covers new models from the AA free tier directly.

Tests

New observation-readers.test.ts (all hermetic — mocked fetchImpl):
fetch + Anthropic enrichment, cache-dir creation, cache-first reuse (no second
fetch), and empty-key short-circuit (no fetch, []).

Full src/tuner/__tests__/wisecron/ suite: 403 pass / 0 fail (run with the AA
key unset to match CI). Biome clean on the touched files.

[Nibbler1250]

Closing this. On review, it wires the injectable benchmarkProvider seam, but its only consumer — proposeEvidencePatch on the proactive EvidenceDrivenSubject face — isn't driven by any scheduler cycle yet (both ProposalEngine.runCycle and the legacy engine call the reactive collectObservations → detectProblems → proposeChange path). So injecting the feeder is a no-op today, and the title oversells it as a behavioural fix.

Rather than ship a node without a live consumer, I'd rather wait until the proactive evidence pass is actually driven by the scheduler (the Subject 2/3 → 3/3 staging), at which point wiring the scout feeder is a natural part of that change. Closing to keep the queue honest — happy to revisit then.