Add bounded cache-aware routing for Hugging Face

[KirschBluteX]

Summary

The built-in Hugging Face provider can keep sending large append-only prompts through a route that reports no cache reuse. This adds bounded, session-scoped adaptation while keeping tau_agent provider-neutral.

• Distinguish absent cache telemetry, reported zero, and positive reuse in the existing usage path.
• Ignore cold, short, and non-comparable requests; warm an automatic route once, then evaluate two comparable misses. Candidate discovery is limited to live conversational providers, at most three routes, and at most nine eligible requests.
• Persist routing evidence, attempted routes, evaluation mode, and every automatic transition through the existing session index and CustomEntry mechanisms. Explicit configured or extension-selected routes remain locked; /route automatic resets evaluation.
• Surface every automatic route change through typed print, transcript, JSON, and TUI events.
• Retain the current usable route when candidate discovery, route switching, or persistence cannot complete.

The policy observes provider-reported cache telemetry only and remains deliberately bounded. It does not change explicit user route precedence or add Hugging Face behavior to the reusable agent harness.

Closes #572

Validation

• uv run pytest -q: 1547 passed, 3 platform-specific tests skipped
• Focused cache-routing suite: 18 passed
• uv run ruff check .
• uv run ruff format --check .
• uv run mypy
• Hugo documentation build and Pagefind indexing
• git diff --check origin/main...HEAD

No dependencies, versions, or generated artifacts changed.