fix: exempt sub-agent sessions from cache warming

[BYK]

Summary

• Sub-agent sessions (OpenCode explore/general agents) are ephemeral (1-3 turns) — cache warming is wasted work for them
• Adds explicit isSubagent flag on SessionState, set from x-parent-session-id header, and checks it in both the idle loop (early skip before expensive work) and shouldWarm() (belt-and-suspenders)
• Also adds early messageCount check in the idle loop to skip all short sessions before profile/histogram computation — a general optimization that benefits non-sub-agent short sessions too

Approach

Two-layer exemption:

1. Explicit isSubagent flag — set on SessionState when x-parent-session-id is detected in the request headers during handleConversationTurn(). Checked in shouldWarm() to block warming unconditionally, even with /keep force.

2. Early idle loop skip — both isSubagent and messageCount < MIN_TURNS_FOR_WARMING * 2 are checked before loadGlobalHistograms(), resolveProfile(), and blendedHistogramForSession(), avoiding wasted computation. Previously, shouldWarm() rejected these sessions but only after all the expensive prep work.

Files changed

File	Change
packages/gateway/src/translate/types.ts	Add isSubagent?: boolean to SessionState
packages/gateway/src/pipeline.ts	Set isSubagent from x-parent-session-id header
packages/gateway/src/idle.ts	Early skip for sub-agent + low-turn sessions in warmup loop
packages/gateway/src/cache-warmer.ts	isSubagent guard in shouldWarm()
packages/gateway/test/cache-warmer.test.ts	2 new tests for sub-agent exemption (normal + /keep)
packages/gateway/test/helpers/idle-worker.ts	Add MIN_TURNS_FOR_WARMING to cache-warmer mock