switchyard_route: avoid Strong/Weak thrashing when Switchyard dies mid-session

[yehuditkerido]

Context

Parent: #2 · POC: #16

The POC in #16 classifies per request (Context::default() each time). If the judge fails mid-chat, on_failure: open continues without a sticky tier and may leave the request unrouted. That can flap Weak ↔ Strong across turns, or change behavior quietly when the judge is flaky.

Epic #2 already calls out thinking through classifier reachability and avoiding turn-by-turn thrashing.

Goal

Define and demonstrate a mid-session failure policy that does not thrash tiers.

Proposed work

• Document current behavior (judge timeout / 5xx / unreachable mid-chat).
• Add a demo scenario (and/or tests) that kills or breaks the judge mid-conversation and shows resulting tiers.
• Implement a sticky policy, e.g.:
  • keep last successful tier for the session key when the judge fails, or
  • fail-closed / reject instead of re-deciding, or
  • another explicit operator-chosen fallback (must be documented).
• Make the chosen policy observable via filter metadata/logs.

Out of scope (unless we decide otherwise)

• Durable cross-replica session store (see historical Pluggable durable session-floor store for switchyard_route (Redis/KV adapter) #3 / enhancements track).
• Broader healthy-path “no downgrade” ratchet (can be a sibling issue).

Acceptance criteria

• Written behavior for “Switchyard unreachable mid-session” under on_failure: open and closed.
• Demo or automated test proves we do not thrash Weak ↔ Strong across turns when the judge dies after a successful Strong (or Weak) decision.
• Metadata/logs make the sticky/fallback path obvious.

[yehuditkerido]

/assign