chore(session): raise arbitration threshold 0.65 → 0.8

[hijzy]

Summary

• Tighten the trust gate for the LLM's new_task verdict in apps/memos-local-plugin/core/session/relation-classifier.ts — bump ARBITRATION_THRESHOLD from 0.65 to 0.8.
• Real-world traces show the primary relation classifier often returns new_task at 0.65–0.75 confidence for messages that are actually sub-tasks of the same project (e.g. "那数据库怎么配" right after "配置Nginx"). The old 0.65 cut-off let those slip through without a second-pass review, falsely splitting one logical task into two.
• Routing more borderline new_task predictions through the bias-toward-follow_up arbitration prompt costs one extra LLM call on those turns, but reduces false topic splits (each false split prematurely closes the episode and leaves a stale lastEpisodeBySession entry that the L2/L3/skill chain has to recover from later).

Behaviour change

Only one knob changes; nothing about the public API, schema, or storage layout.

Path	Before	After
core/session/relation-classifier.ts::ARBITRATION_THRESHOLD	0.65	0.8

Decision flow (unchanged structure, only the cut-off moves):

1. LLM returns new_task with confidence < ARBITRATION_THRESHOLD → run the second-pass arbitration prompt (biased toward follow_up).
2. Otherwise → take the LLM verdict at face value.

Test plan

• npx vitest run tests/unit/session/relation-classifier.test.ts — 16 tests passed (existing arbitration test uses a 0.5 confidence input which is below both old and new thresholds, so it still exercises the same path).
• Smoke check on a real session where consecutive sub-task messages were previously being mis-split — confirm relation.classified log now shows signals: ["llm", "arbitration_override"] for those turns and the episode stays open.