Thinking router re-fires on every timeout/auth resume slice

[sentry-junior]

selectTurnThinkingLevel makes a Haiku LLM call to pick a complexity level for the upcoming turn. It is called unconditionally every time generateAssistantReply runs — including on every timeout-resume and auth-resume slice, where the turn context is identical to the original invocation.

Root cause: respond.ts:843 has no resumedFromSessionRecord guard, and AgentTurnSessionRecord does not persist thinkingSelection, so resumed slices can't skip or reuse the prior result.

Impact: each resume slice pays an extra Haiku round-trip for a classification that will return the same answer (temperature: 0, same input). With AGENT_TURN_TIMEOUT_RESUME_MAX_SLICES = 48, a long sandbox task can generate 10–48× the expected router calls per turn.

Affected files:

• packages/junior/src/chat/respond.ts — unconditional selectTurnThinkingLevel call at line 843
• packages/junior/src/chat/state/turn-session.ts — AgentTurnSessionRecord has no thinkingSelection field
• packages/junior/src/chat/services/turn-session-record.ts — persistRunningSessionRecord / persistAuthPauseSessionRecord / persistTimeoutSessionRecord need to carry the selection forward

Proposed fix:

• Add an optional thinkingSelection field to AgentTurnSessionRecord and its storage/parse path (backward-compatible, fall back to recompute if absent)
• Persist the selection at the same time the first safe Pi boundary is written
• Skip selectTurnThinkingLevel on resume when a persisted selection is available, adding a "app.ai.thinking_level_source": "session_record" | "router" span attr for traceability

See also: #431 (compaction rework, related session-record work)

Action taken on behalf of David Cramer.