feat(libsy): re-decide the classifier tier on fresh user turns

[gburachas]

Problem

llm_classifier classifies a conversation once and, with session_affinity, holds that decision forever. Both halves are wrong for long agentic sessions:

• Classify every turn and a tool chain switches tier mid-flight — the weak tier picks up a conversation the strong tier was three tool calls into, with no idea what it was doing.
• Classify once (session_affinity) and the opening verdict goes stale. A session that opened "summarise this file" becomes a hard debugging session ten turns later and is still served by the tier chosen for the summary.

The unit that should drive a routing decision is not the request and not the session — it is the user turn. Tool results are the agent continuing work the user already asked for.

Proposal

A followup table that re-classifies when the user asks something new and pins that target across the tool-call turns in between:

[routes.assistant]
type = "llm_classifier"
followup = { refresh = "on_user_turn", recent_turn_window = 4 }

Key	Default	Meaning
refresh	on_user_turn	on_user_turn re-classifies on each new user message; pin_once holds the first decision
recent_turn_window	unset	Trailing turns the judge sees. At least 1
fail_open_tier	capable	Tier served when the judge yields nothing usable and nothing is pinned. Mode capability only
abstain_selector	unset	JSON Pointer to a boolean abstain flag in the verdict. Mode custom only

Available on modes capability and custom. Rejected alongside session_affinity, whose latch is first-wins and would prevent every refresh — silently turning on_user_turn into pin_once; use refresh = "pin_once" if that is what you want.

The deliberate asymmetry, which is the part worth reviewing:

• an unusable verdict, or an unreachable judge, holds the pinned target. No evidence is not evidence of change. It falls open only when nothing is pinned.
• in mode custom, an explicit abstain outranks the pin — the judge declining to decide is itself a decision. Mode capability has no equivalent, because its packaged rubric answers with a solve probability and carries no abstain field.
• neither path retries the judge.

Relationship to #298 and #308

These overlap and the difference is intentional. #298's RoutingIdentity and #308's Harbor proxy_x_session_id normalisation both improve who a conversation is; this is about when its tier is re-decided. The pin lives in per-session State rather than the affinity map precisely because AffinityRouter's assignment is first-wins and process-local, so it can never be refreshed — which is the behaviour this issue exists to change.

Where a request carries no session id the pin has nothing to persist in, so it can optionally be keyed on a hash of the first user message, the same fallback session_affinity already offers.

Rebased onto current main and green. Depends on nothing else; sibling issue for the graded-verdict labels map filed separately.