Simplify memory core and protect snapshots and concurrent reads

[Ninadnj]

Returned Python entries previously exposed live store state: changing a result could bypass revisions and leave text inconsistent with its search vector. Concurrent reads could also observe a change whose disk save later failed.

This candidate makes returned entries independent snapshots and serializes operations within each store object. Record rules and JSON persistence now have explicit modules; MemoryStore continues to own memory operations, retrieval and transactions. Existing format-3 data and IDs remain compatible.

Other changes:

• Keep startup handoff selection and recall on one snapshot during other writers' corrections.
• Treat literal tokenizer markers as ordinary memory text, including statistics and budgeted MCP output.
• Default new stores to offline hashing, preserve saved backend configuration, and report invalid choices.
• Replace dedup_threshold with deduplicate=False; make agent keyword-only in the Python write APIs.
• Share rendering tags, avoid copying histories just to count entries, and add consistent lint/format checks.
• Document architecture, write/read flow, tradeoffs, migration and verification for v0.4.0rc2.

Validation on macOS / Python 3.12:

• 330 tests passed with MCP 2.2.0; only the optional semantic-model module skipped.
• 304 tests passed in the minimal environment, with 9 optional checks skipped (including exact-tokenizer cases).
• 89 focused tests passed with MCP 1.30.0, including the real two-process demo.
• All 30 task fixtures validated; published retrieval JSON/Markdown reproduced exactly.
• Wheel/sdist build, fresh wheel installation, lint, formatting and documentation links passed.

The new regression suite reproduced failures before implementation. Retrieval quality is unchanged; semantic-model and live-agent performance remain unverified. See docs/migration-v0.4.md before upgrading Python callers.

The pre-merge fixes add ten regression cases (9 failed before the fixes, all 10 now pass), covering cross-instance startup consistency, literal markers, persisted-memory budgets, and MCP roundtrips.