feat(memory): optional Dongtian memory retrieval for system prompt context

[siaochuan]

Closes #67

Summary

Add an optional integration layer that retrieves relevant conversation snippets from a local Dongtian memory database and injects them into the system prompt. Default behavior is completely unchanged — the feature is off unless explicitly enabled.

How it works

User sends a message
  → OH searches the local Dongtian palace DB (FTS5 keyword match)
  → Top-N relevant snippets injected into system prompt as context
  → Agent sees historical conversation context from past sessions

Configuration

{
  "memory": {
    "enabled": true,
    "dongtian_enabled": true,
    "dongtian_db_path": "~/.dongtian/palace.db",
    "dongtian_limit": 5,
    "dongtian_max_chars": 1200
  }
}

All fields have safe defaults. Set `dongtian_enabled: true` to activate.

Design decisions

• Off by default — zero behavior change for existing installs
• Zero dependencies — uses only Python stdlib `sqlite3`; no `dongtian` package import needed
• Read-only — opens the DB in `?mode=ro` URI mode; never writes
• FTS5 only — no embedding API calls, no network, works fully offline
• Timeout protected — 200ms SQLite timeout, graceful fallback to empty on any error
• Injection safety — snippets are wrapped with an untrusted-content warning telling the model not to follow instructions inside retrieved memories
• Chinese-aware tokenizer — FTS query builder handles CJK characters

Relation to #67

This addresses the same need as #67 (external memory retrieval for system prompt) with a different backend:

	PowerMem (#67)	Dongtian (this PR)
Backend	HTTP API server	Local SQLite file
Dependencies	Running server + SDK	None (stdlib sqlite3)
Search	Vector + full-text + graph	FTS5 keyword (+ optional embeddings via MCP)
Setup	Install + configure server	Just point to a `.db` file
Offline	No	Yes

Both approaches are valid for different use cases. A future `MemoryProvider` protocol could unify them — see discussion on #67.

Files

File	Purpose
`memory/dongtian.py`	FTS5 search against palace DB (166 lines)
`config/settings.py`	`MemorySettings` fields for dongtian config
`prompts/context.py`	Snippet injection into system prompt
`tests/test_prompts/test_dongtian_memory.py`	End-to-end tests (creates DB, searches, verifies injection)

What is Dongtian?

Dongtian is a lightweight AI conversation memory system (~1200 lines, SQLite + FTS5 + optional embeddings). It ingests sessions from Claude Code, Codex, OpenCode/DeepSeek, ChatGPT, and Slack into a structured searchable database. It also runs as an MCP server with 15 tools for search, ingestion, knowledge graph, and SSH remote sync.