Closes #4732

Problem

Configuring a planner_model forces every conversation through the planner, even for simple questions like "who wrote this file?" or "is this safe?". The TaskWarrantsPlanner heuristic only bypassed the planner for inputs starting with what/why/how/show/run + 5 Chinese prefixes — everything else (including who, where, when, explain, describe, is, are, can, does, etc.) was sent to the planner unnecessarily, doubling latency.

Additionally, the existing LLM AutoPlanClassifier (which makes a tiny 80-token classification call) was only wired into the single-model shouldAutoPlan path — the two-model Coordinator never used it.

Changes

1. Widened isLowRiskQuestion (internal/control/auto_plan.go)

Added 20+ missing English prefixes: who, where, when, which, whose, whom, explain, describe, tell, list, summarize, summarise, compare, difference, is, are, can, could, do, does, did, should, would, will, what's (apostrophe variant handled via normalization).

Added 8 missing Chinese prefixes: 介绍一下, 说一下, 帮我看, 帮我查, 是什么, 有没有, 能不能, 可以吗, 对吗, 是不是, 请问.

Complex intent terms (implement, refactor, migrate, etc.) still override these prefixes — "how do I implement X" still goes through the planner.

2. Exported AutoPlanClassifier + added NewPlannerGate (internal/control/auto_plan.go)

Exported the autoPlanClassifier interface as AutoPlanClassifier. Added NewPlannerGate(classifier) that wraps TaskWarrantsPlanner with LLM classifier support for borderline cases (score ≤ 2), mirroring the single-model shouldAutoPlan path. When classifier is nil, falls back to pure heuristic.

3. Wired classifier into Coordinator (internal/boot/boot.go)

Moved classifier creation BEFORE the Coordinator creation. Changed control.TaskWarrantsPlanner → control.NewPlannerGate(classifier) so the two-model path gets the same LLM classification the single-model path already had.

4. Updated controller types (internal/control/controller.go)

Changed Classifier field from unexported autoPlanClassifier to exported AutoPlanClassifier.

Tests

• 35 test cases for TaskWarrantsPlanner (was 12) — covers all new English + Chinese prefixes + complex intent overrides
• 3 new tests for NewPlannerGate: nil classifier fallback, classifier override (both true/false), low-risk skip

Files Changed