feat(memory): add global and session memory search

[xiami762]

Summary

• Add persistent Memory and Session transcript search through the existing memory_search tool.
• Keep embeddings opt-in while providing indexed keyword search, incremental reconciliation, and citations.

Key Changes

• Add hybrid Memory search across file-backed Memory and persisted Session history, with source filtering and citation decoration.
• Add incremental Memory file indexing, Session transcript reconciliation, chunking, embedding caching, and stale-index cleanup.
• Persist searchable Session message and tool-part data and add the required SQLite storage helpers.
• Add memory.search.embedding.enabled configuration, defaulting to false, and write the minimal Memory Search configuration on first run.
• Add focused coverage for Memory scope handling, Session transcript search, configuration initialization, storage, and message-part persistence.

Impact Scope

• User-visible behavior: Agents can use memory_search to retrieve durable Memory and, when requested, Session history with file/line citations.
• Compatibility / migration: Existing Memory files and Sessions remain valid; search tables are created through the existing storage initialization path.
• Configuration / environment: Adds optional memory.search.embedding settings. Embeddings are disabled by default and require no new secret unless enabled.
• Dependencies: No new third-party dependencies.
• Performance / resources: Adds incremental indexing and searchable Session storage. Embedding cost and network usage occur only when embeddings are enabled.
• Security / permissions: Session transcripts and tool traces become searchable data. Review project/session ownership filters and the set of Memory paths included in each search scope.

Business Logic to Review

• MemoryManager reconciliation and the boundary between Global Memory, Project Memory, Daily Memory, and Session history.
• MemoryIndexer file discovery, project scoping, stale-file cleanup, and idempotent incremental updates.
• Session transcript watermarks, tool-part serialization/truncation, deleted-message handling, and concurrent reconciliation.
• Vector/FTS result merging, score normalization, source filters, and citation line ranges.

Why This Approach

• Reuses the existing file-backed Memory model and memory_search interface instead of introducing a separate Memory CRUD tool.
• Uses SQLite-backed incremental indexes so search remains available without embeddings and can add semantic recall when explicitly enabled.

Test Plan

• uv run pytest tests/memory/test_memory_scope.py tests/memory/test_session_transcript_search.py tests/config/test_config_init.py tests/storage/test_storage.py tests/session/test_message_parts_persistence.py -q (94 passed)

Compatibility, Migration & Rollback

• No manual migration is required; storage initialization creates the additional search structures.
• Embeddings remain off by default. Roll back by reverting this PR; existing Memory Markdown files are unaffected.