Replace query() with ClaudeSDKClient for native session management

[RichardAtCT]

Summary

• Swap stateless query() for per-request ClaudeSDKClient instances, eliminating temp session IDs (temp_*) and the delete-old/save-new session swap pattern
• New sessions now created with session_id="" and deferred storage save until Claude responds with a real ID
• update_session() takes a ClaudeSession object directly instead of a string ID, simplifying the facade's post-execution flow
• Uses ResultMessage.result for content extraction with fallback to message parsing

Test plan

• All 379 tests pass (poetry run pytest tests/ -v)
• No new flake8 or mypy errors introduced
• New sessions get real session_id from ResultMessage (not temp)
• Resumed sessions pass options.resume correctly to ClaudeSDKClient
• ResultMessage.result used when available, falls back to message extraction
• Manual smoke test with live Telegram bot

🤖 Generated with Claude Code

[RichardAtCT]

PR Review
Reviewed head: fe6928a07575bdcc2ff580889de586e5cf5e0871

Summary

• Migrates from query() to per-request ClaudeSDKClient, which is a cleaner lifecycle model.
• Removes temp session IDs and defers persistence until a real Claude session ID is returned.
• Adds strong unit coverage around resume behavior, result extraction, timeout/error paths, and MCP option passing.

What looks good

• src/claude/sdk_integration.py: using ResultMessage.result first with a fallback is the right extraction order.
• tests/unit/test_claude/test_sdk_integration.py: coverage improved substantially and tracks intent of the refactor.
• src/claude/facade.py: simplification from temp-ID checks to truthy real-ID checks is much easier to reason about.

Issues / questions

1. [Blocker] src/claude/session.py (update_session) — if Claude returns no response.session_id for a new session, we still set is_new_session=False and then skip save_session because session.session_id is empty. That leaves an untracked in-memory session that cannot be resumed later and can silently lose continuity after process restart. Can we either (a) fail hard for new sessions without an ID, or (b) persist a recoverable placeholder state and retry ID binding on next turn?

2. [Important] src/claude/facade.py + src/claude/session.py — after update_session, we force response.session_id = session.session_id. If session.session_id is still empty, downstream callers/logging get an empty ID that looks “successful” but is unusable. Suggest explicit guard + warning/error path when final session ID is empty.

3. [Important] tests/unit/test_claude/test_session.py (test_session_limit_enforcement) — assertion expects len(persisted) == 1 after creating a third session, but comment says user should still have 2 persisted sessions. The test currently codifies behavior where new sessions are unsaved pre-response, but the mismatch between comment/expectation makes intent unclear. Worth tightening to avoid future regressions in retention logic.

Suggested tests (if needed)

• Add a unit test for update_session(new_session, response_without_id) that verifies explicit behavior (raise, warning + persisted placeholder, or other chosen policy).
• Add facade-level test ensuring empty final session_id is surfaced as an error/diagnostic, not silently propagated.
• Add an integration-style test: create new session -> execute -> restart simulation -> resume, to validate persistence semantics end-to-end.

Verdict

• ⚠️ Merge after fixes