Add file/Redis/SQL conversation stores, modernize toolchain

[HartBrook]

Summary

Implements the three conversation store backends that llmpane already advertised but never shipped, and brings the build, lint, and type-checking stack back up to date.

The bug this fixes

create_store() accepted "file", "redis", and "sql". pyproject.toml offered matching [file], [redis], [sql], and [stores] extras. The README documented the calls verbatim:

store = create_store("redis", url="redis://localhost:6379")

None of those modules existed. Every one of those paths raised ModuleNotFoundError, so anyone following the README hit a wall on first use. create_store had no test coverage, which is why CI stayed green over it.

New store backends

All three implement the existing ConversationStore protocol and are drop-in interchangeable with InMemoryStore:

Backend	Extra	Notes
FileStore	llmpane[file]	One JSON document per conversation. Atomic write-then-rename, so an interrupted write can't truncate a good file.
RedisStore	llmpane[redis]	JSON values plus a sorted set keyed by updated_at, so list_conversations() paginates without scanning the keyspace. Supports ttl.
SQLStore	llmpane[sql]	SQLAlchemy 2.0 async (SQLite, PostgreSQL, ...). Timestamps mirrored into columns so ordering and pagination happen in the database.

A couple of deliberate design points:

• FileStore validates conversation IDs. They become filenames, so they're checked against a strict allowlist and anything else raises ValueError. There are explicit path-traversal tests.
• RedisStore and SQLStore don't close what they don't own. Pass your own client= / engine= and it survives close(); only a connection llmpane created gets torn down.
• Failure modes are legible. An unknown backend name lists the valid ones; a missing optional dependency names the extra that provides it, rather than surfacing a raw import error.

Tests are a shared conformance suite parameterized across all four backends, so the contract is verified identically for each, plus dedicated coverage for path safety and connection ownership. Python tests go from 122 to 202.

⚠️ Breaking change

The minimum supported Python is now 3.12 (previously 3.10). Python 3.10 and 3.11 are no longer tested or supported. This lands before any published release, so no released version changes behaviour.

The React package's supported peer range is unchanged (React >= 18).

Toolchain modernization

mypy --strict was configured but never actually run, and had accumulated 29 errors. Those are fixed, and mypy, ruff format --check, and linting of the tests/ directory are now CI gates rather than aspirational config.

Lint and test dependencies were several majors behind:

	Before	After
ESLint	8 (end of life, .eslintrc.cjs)	9, flat config
typescript-eslint	6	8
Vitest	1	4
jsdom	22	30
@vitejs/plugin-react	4	6
CI actions	v4	v7 / v8 / v9
Node (CI)	20 (EOL)	22 and 24
Python (CI)	3.10–3.12	3.12–3.14

Three upgrades were deliberately not taken, each for a concrete reason rather than caution:

• ESLint 9 rather than 10 — eslint-plugin-react has no ESLint 10 release, so moving to 10 would mean dropping React JSX lint coverage entirely.
• TypeScript 5.9 rather than 7 — typescript-eslint@8 declares typescript: ">=4.8.4 <6.1.0", so TS 7 isn't lintable yet.
• UP046 (PEP 695 generics) ignored, with the reason recorded in the config — it scopes type parameters to a single class, but MetadataT is intentionally shared across modules so ChatMessage and Conversation stay parameterized by the same type.

The newer react-hooks rules surfaced three findings. One was a genuine bug — a ref written during render in useConversations — now synced in an effect. The other two are suppressed with written justifications rather than reshaping working code to satisfy a conservative analysis.

Verification

• Python: 202 passed, mypy --strict clean, ruff and format clean. Run against 3.12, 3.13, and 3.14.
• React: 130 passed, lint clean, typecheck clean, build OK. Run against Node 22 and 24.
• Downstream: validated against a production consumer of llmpane without modifying it, by shadowing its vendored copy at import time. Its full llmpane-dependent suite passed unchanged, and a custom ConversationStore implementation still satisfies the re-annotated protocol.

That downstream check paid for itself: it caught a cast("CursorResult[Any]", ...) that was fragile across SQLAlchemy 2.0.x patch releases, which type Session.execute() differently. Now version-robust.

Commits are scoped by area (Python / React / CI) to make review easier. CI is green on all five matrix jobs.

Follow-ups, intentionally not in this PR

Reviewing the codebase turned up four issues that are behaviour and security changes rather than modernization. They're listed here so they aren't lost, and are best handled as separate PRs:

1. Error detail leakage. errors.py attaches details={"original_error": str(exc)} to every classified error, and that reaches the client. Provider exceptions can carry request content or response bodies.
2. Terminal message ID mismatch. The done=True chunk carries a different ID than the one store.add_message persists, so clients can hold an ID that doesn't exist server-side. The terminal is also emitted before persistence, so a persistence failure still reports success.
3. Conversation ID trust. An unknown caller-supplied conversation ID is created under that same ID.
4. Protocol parity. Python's StreamChunk sends usage: ModelUsage, but the React types have no corresponding field.

Items 1 and 2 are the ones worth prioritizing.