fix(acp): persist thread→session mapping to survive pod restarts

[JARVIS-coding-Agent]

Context

Closes #456

After a pod restart, all Discord thread-to-session mappings are lost because the suspended HashMap in SessionPool is purely in-memory. Even though kiro-cli correctly persists session state to PVC, OpenAB cannot associate returning threads with their previous sessions — every thread gets a brand-new session instead of resuming.

Summary

Persist the thread_id → session_id mapping (suspended HashMap) to a JSON file on PVC so it survives pod restarts. On startup, the mapping is reloaded and threads can resume their previous sessions via session/load.

Changes

• Added mapping_path: PathBuf field to SessionPool, derived from config.working_dir
• load_mapping(): reads thread_map.json on startup; falls back to empty map if file is missing or corrupt
• save_mapping(): writes the suspended map to disk after every mutation (3 save points: get_or_create, cleanup_idle, shutdown)
• shutdown() now persists active sessions into the mapping before clearing, so graceful restarts preserve all thread associations

Design Decisions

• JSON file on PVC — simplest approach that matches the existing persistence model (kiro-cli already writes session JSON to PVC). No new dependencies needed (serde_json already in Cargo.toml).
• Save on every mutation — ensures mapping is always up-to-date even on ungraceful shutdown. The write frequency is low (only on session eviction, idle cleanup, or shutdown) so disk I/O is negligible.
• Graceful degradation — corrupt or missing thread_map.json logs a warning and starts fresh instead of crashing. Failed writes also warn without panicking.
• No cleanup of stale entries — out of scope for this fix. Stale mappings pointing to deleted session files are harmless: session/load will fail and fall back to session/new.

Discord Discussion URL: https://discord.com/channels/1491295327620169908/1495245278205182093

[JARVIS-coding-Agent]

Review Feedback 處理狀態

🟡 NIT — 全部修了

#	Feedback	狀態
1	unwrap_or_default() 靜默寫空檔	✅ 改成 match，序列化失敗 log warning + early return
2	&PathBuf → &Path	✅ 改了
3	to_string → to_string_pretty	✅ 改了

🔴 SUGGESTED CHANGES

#	Feedback	狀態	說明
1	Atomic write	✅ 改了	先寫 .json.tmp 再 fs::rename，同 filesystem 上 rename 是 atomic
2	shutdown() 的 try_lock() skip 問題	✅ 改了	改用 .lock().await，shutdown 路徑確保所有 active session 都存到
3	Stale entry 清理	❌ 不改	mapping 只存兩個短字串（thread_id → session_id），1000 條也才幾 KB。加 pruning 需要知道 kiro-cli session file 路徑格式，會增加跨模組耦合，不適合在這個 bug fix 裡做。可以開 follow-up issue