fix(session): tighten empty-step guard so genuine empty tool calls always halt

[wqymi]

Summary

• Add unconditional empty-step guard in processor.ts BEFORE the doom-loop check: genuine empty tool calls (undefined/null/empty object, even with _meta fields) are always counted, regardless of error/summary/structured/finish state
• When emptyStepCount hits EMPTY_STEP_THRESHOLD=3, throws to halt the stream and sets assistant.error

Changes

• packages/opencode/src/session/processor.ts: Added EMPTY_STEP_THRESHOLD = 3, isEmptyInput() helper (filters _meta-prefixed keys), emptyStepCount to ProcessorContext, and the guard logic in the tool-call handler
• packages/opencode/test/session/processor-effect.test.ts: Regression test — single LLM response with 3 consecutive empty tool calls (one with {_meta:'harmless'}), asserts processor returns 'stop' and assistant.error is set

Why

The doom-loop check only catches last-3 SAME tool name + identical serialized input. The prompt.ts:2734-2742 passthrough returns 'none' (skipping empty detection) when assistant has error/summary/structured or finish=content-filter/error. Genuine empty tool calls slip through both guards.

Verification

• bun typecheck EXIT0
• bun test test/session/processor-effect.test.ts — 12 pass, 1 skip (pre-existing), 0 fail