feat(honcho): optional lightweight cached startup context for recallMode=tools

[teyrebaz33]

Closes #2150

What

Adds an opt-in background prewarm for recallMode=tools that fetches a small user snapshot (representation + peer card) from Honcho on /new or session reset, and injects it into the first real turn's user message.

How

• HonchoClientConfig: new tools_startup_context: bool = False field (JSON key: toolsStartupContext)
• HonchoSessionManager.fetch_startup_snapshot(peer_id): fetches user representation + peer card without requiring an active session in local cache
• run_agent.py: _schedule_honcho_startup_prewarm() fires a daemon thread on session init; _pop_startup_snapshot() consumes it once on first turn
• cli.py: prewarm triggered in new_session()
• gateway/run.py: prewarm triggered after reset, _honcho_startup_cache keyed by peer_name

Design decisions

• Injected via _honcho_turn_context (user message append), not _honcho_context (system prompt) — prompt cache prefix stays stable
• pop() semantics — consumed exactly once, subsequent turns unaffected
• not conversation_history guard — first turn only
• Keyed by peer_name (stable Honcho identity), not ephemeral session ID
• V1 scope: representation + peer card only. No dialectic, no AI peer context, no continuity recall

Not included (by design)

Continuity-style recall is intentionally excluded per issue spec — /new semantically means a fresh session.

[gaelcado]

Whoops didn't catch that ! I implemented my version, have you had any success testing yours ?

[teknium1]

Thanks for this contribution. Closing because the architecture it targets no longer exists: honcho_integration/ was removed when honcho became a memory-provider plugin at plugins/memory/honcho/, and current policy requires memory-provider changes to live entirely inside the plugin directory (this PR wires through cli.py, gateway/run.py, and run_agent.py, which the plugin surface now forbids).

The ideas here are still good — the current plugin already has prewarm infrastructure for context/hybrid recall modes, but tools mode defers init with no cached startup snapshot. If you'd like to revive this, a plugin-only implementation using the existing prewarm/_prefetch_result machinery in plugins/memory/honcho/__init__.py would be welcome (pop-once semantics and user-message injection to preserve prompt caching, as you designed).