chore(#47/#48): retry classification, missing failover tests, type-check tests

[arniesaha]

Summary

Closes #47 and #48 in one PR (both touch the same testing/observability surface area introduced by #45/#46).

#47 — failover follow-ups

• Retry classification: isRetryableDownstreamError now consults err.cause.code authoritatively (undici wraps the real network error there) for ECONNREFUSED, ECONNRESET, ENOTFOUND, EAI_AGAIN, ETIMEDOUT, EPIPE, and UND_ERR_*. The "fetch failed" message-substring check is kept as a fallback so we don't regress current behavior if Node ever stops attaching cause.
• Three new tests:
  • parameterized cause.code classification (6 codes + AbortError + plain non-network error)
  • 3-provider chain-order: a → b → c, primary and first fallback error, third succeeds, asserts call order
  • streaming headers-sent guard with a retryable mid-stream error: provider A streams a frame (flips headersSent), then errors with ECONNRESET; dispatcher must not failover (would corrupt SSE), provider B is never called
• Enriched mux.failover_hop log: now includes fromUrl/toUrl (looked up from config.providers) and a reason field that surfaces cause.code on network errors (e.g. TypeError:ECONNRESET)

#48 — type-check tests

• tsconfig.test.json extends the base config with tests/**/*.ts included
• npm run check now runs tsc --noEmit against both tsconfig.json and tsconfig.test.json
• Fixed the drift this surfaced: shared route fixture in tests/downstream.test.ts was missing providerId/fallbackProviderIds, and provider mocks in tests/policy.test.ts were missing protocols

Deferred from #47

Item 3 (rename prov.route.provider_id → prov.effective.provider_id to keep the original route decision immutable) intentionally NOT in this PR — needs coordination with the AgentWeave Routing tab (arniesaha/agentweave#164) before flipping the schema.

Testing

• npm test → 118/118 (was 107)
• npm run check → clean (src + tests)