fix(writer): explicit timeout + retries on every LLM call

[Chen17-sq]

Summary — silent reducer / classifier hang

writer/llm.call_llm calls litellm.completion(**kwargs) with no timeout and no num_retries. Production litellm versions default the request to "wait forever", so a stuck connection (slow provider, partial response, dead TCP socket) blocks the reducer or classifier daemon thread indefinitely. The daemon stays "alive" but no durable facts get written and nothing in the logs alerts the user. Combined with the orphan-session recovery in #18, this is the last hang point that prevents end-to-end self-healing.

Fix

Two module-level defaults in writer/llm.py:

• DEFAULT_TIMEOUT_SECONDS = 120 — generous for a long session reduce without being absurd; reasoning-model / slow local-model users can raise it per stage.
• DEFAULT_NUM_RETRIES = 2 — 3 total attempts. litellm uses tenacity-style backoff and respects Retry-After headers, so this absorbs a brief 429 / 502 blip without adding noticeable latency.

Two new fields on ModelConfig: timeout_seconds: int | None = None and num_retries: int | None = None. None means "use module default", so existing configs that don't mention these parse cleanly and get the safe defaults — no breakage for upgrades.

The default config template gains commented-out lines documenting both knobs:

[models.default]
# timeout_seconds = 120 # per-call timeout; raise this if you run a slow local model
# num_retries = 2       # automatic retries on transient errors (429, network blips)

Worst-case behavior

A fully unresponsive provider keeps the reducer thread blocked for timeout * (1 + num_retries) = 360 s before the exception propagates to the reducer's outer except, which marks the session failed and lets the safety-net retry tick handle it. Far better than today's unbounded wait.

Relationship to other PRs

• fix(session): recover orphan active sessions on daemon startup #18 (orphan-session recovery) — orthogonal but complementary. fix(session): recover orphan active sessions on daemon startup #18 ensures sessions don't get stuck in active after a hard crash; this PR ensures the eventual reduce of those sessions can't hang on a slow LLM provider. Together: complete self-healing.
• fix(store): atomic writes for memory files #11 / fix(store): per-path lock for concurrent memory-file writes #12 — different files, no overlap.

Mergeable in any order.

Out of scope

• Streaming / partial-response timeouts — OpenChronicle uses non-streaming completions, so request timeout is total response time. Streaming would need separate inactivity-timeout handling.
• Sanity-clamping timeout_seconds = 0 or num_retries = 100 — explicit user choices are passed through; clamping would silently override what they wrote.

Test plan

• uv run pytest — 76/76 pass
• uv run ruff check src/ tests/test_writer_llm.py tests/test_config.py — clean

New tests cover both layers:

tests/test_writer_llm.py:

• test_call_llm_passes_default_timeout_and_retries — patches litellm.completion, asserts module defaults are forwarded.
• test_call_llm_per_stage_timeout_override — stage-level override wins.
• test_call_llm_other_stages_keep_defaults_when_one_is_overridden — sibling stages don't bleed.
• test_call_llm_mock_path_bypasses_litellm — OPENCHRONICLE_LLM_MOCK=1 still short-circuits before litellm import.

tests/test_config.py:

• test_timeout_and_retries_inherit_from_default — TOML inheritance from [models.default] to a stage section.
• test_timeout_per_stage_overrides_default — per-stage override wins over default.
• test_missing_timeout_defaults_to_none — old configs without the fields parse as None, triggering the "use module default" path in the wrapper.

[gemini-code-assist]

Code Review

This pull request introduces configurable timeouts and retry logic for LLM calls to prevent indefinite blocking of daemon threads. It adds timeout_seconds and num_retries fields to the ModelConfig dataclass, establishes module-level defaults (120 seconds and 2 retries, respectively), and ensures these parameters are correctly passed to the litellm library. Comprehensive tests were added to verify configuration inheritance, overrides, and correct parameter forwarding. I have no feedback to provide.