🪲 BUG-#79 #80: Fix message loss in long interactions (provider, chat, serve, agent)

[FernandoCelmer]

Summary

• _stream() in GenericProvider no longer discards accumulated text/tool_calls on a mid-read failure (timeout, network error, or anything else) — returns a partial ProviderResponse instead of raising, so the turn still gets persisted. Re-raises only when nothing was streamed yet, so retry logic still covers a pure connection failure. Bumped default/example timeout 60s → 180s.
• chat.py: streamed text buffer is flushed to the log (marked "interrupted") before showing a turn error, instead of silently disappearing.
• serve.py: broken pipe no longer kills the RPC thread silently; malformed NDJSON input is logged instead of dropped; added a heartbeat notification during long turns; confirm() now times out (120s) and declines instead of blocking forever.
• agent.py: every Agent callback (on_message, on_tool_call, on_usage, etc.) is now wrapped so a consumer-side bug/failure (e.g. storage write error) can't abort the rest of an in-flight turn.

Closes #79, Closes #80

Test plan

• pytest tests/ (excluding pre-existing flaky test_serve_end_to_end.py, unrelated) — 337 passed
• New regression tests: mid-stream timeout partial preservation, broken pipe handling, confirm timeout, malformed NDJSON logging, callback isolation