Name the serving route on advisor errors and failed fallbacks

[judagent]

Follow-up hardening for the PR #11 anthropic→opencode fallback, from live
review of the first fallback-era calls:

• Error rows now record advisorProvider whenever a route was chosen, so a
  failed generation on the fallback no longer looks like an Anthropic
  failure. Rows where no route was chosen still omit it.
• A failed fallback reports both reasons with the primary first, keeping
  classifyAdvisorError on the configured route.
• The fallback support check is wrapped so its own failure can never mask
  the primary error.
• Regression test: the shared advisor session re-pins to opencode on
  fallback and back to anthropic when the primary recovers.

Considered and declined: scoping the fallback to unavailability reasons by
matching reason strings. The guarded-against case (an effort invalid on the
primary but valid on opencode) succeeds on an equivalent route and is now
visible via advisorProvider, so string-based gating would add brittleness
without a real safety win.

Verification

• bun test: 363 pass, 0 fail; tsc --noEmit clean.
• Live smoke on master (pre-fix): advisor_response on
  anthropic/claude-opus-5-5#xhigh with benchmarks matched across the
  board, requester meta/muse-spark-1.3#max via invocation_message.

Models: anthropic/claude-opus-5-5:xhigh, meta/muse-spark-1.3:max, anthropic/claude-fable-5-1:xhigh