Integration

[xmyleader]

Summary
This PR integrates four complementary router improvements into main: quality-aware context routing, cache-aware model switching, HALO health-aware fallback recovery, and a reliable evaluation attempt ledger.
Key Changes

• Adds bounded, context-aware token-saver classification with continuation detection, confidence guards, and routing diagnostics.
• Introduces cache-aware stay-versus-switch cost estimation, pricing provenance, and prompt-cache plan rebuilding.
• Adds HALO recovery with provider health tracking, circuit-breaker behavior, chain-wide retry budgets, deadlines, and safe stream continuation.
• Records judge, primary, retry, fallback, and compaction attempts in an append-only evaluation ledger with usage and cost provenance.
• Adds evaluation scripts, schemas, benchmark reports, and focused regression coverage.
• Preserves correct retry/fallback relationships and prevents duplicate attempt accounting.
  Validation
• npm run build passes.
• 157 focused router, evaluation, streaming, and compaction tests pass.
• All four feature branch histories are included in integration.