strip harmony-style control tokens from local-engine render

[cjus]

Summary

• Local models (gpt-oss variants surfaced via Ollama/LMStudio) leak harmony channel markers like <|channel>thought<channel|> and stray <|start|> / <|end|> / <|message|> / <|return|> tokens into the user-facing Telegram/web reply.
• Added scrubLocalControlTokens(text) in src/local.ts and applied it at all four render touch-points: renderStreamingStub, renderFinal, renderToolLoopStub, renderToolLoopFinal.
• Display-only — audit.response and recentChatTurns history still see the raw model output for forensics. Local engine only — Claude tiers are untouched.

Why

Before: <|channel>thought<channel|>It is 8:44 PM in London. rendered to the user.
After: It is 8:44 PM in London.

Three-pass scrubber:

1. Collapse paired blocks <|name>...<name|> (drops the channel-header content too).
2. Suppress unclosed opener at end of buffer (prevents mid-stream flicker until the closing marker arrives in a later delta).
3. Strip stray symmetric harmony tokens (<|start|>, <|end|>, etc.) and orphan closers.

Test plan

• npm run typecheck clean
• bun test — 766 pass, 0 fail (added 7 new tests under describe("scrubLocalControlTokens", ...) in src/local.test.ts)
• Manual: send a message via Telegram / web UI with LMStudio backend serving a gpt-oss / similar model; confirm channel markers are gone from the rendered reply but raw output is preserved in audit.response