feat(config): per-operation reasoning_effort override (#2998)

[nicoloboschi]

Summary

reasoning_effort was the only LLM request setting with no per-operation override. Retain, reflect and consolidation each already accept their own PROVIDER, API_KEY, MODEL, BASE_URL, MAX_CONCURRENT, MAX_RETRIES, INITIAL_BACKOFF, MAX_BACKOFF, TIMEOUT and LITELLMROUTER_CONFIG — but all four operation configs read the single global HINDSIGHT_API_LLM_REASONING_EFFORT.

When one operation requires a specific value, every other operation is forced to take it too. Concretely: some OpenAI reasoning models reject function tools unless reasoning_effort is exactly "none". Reflect is a tool-calling search loop, so pointing it at such a model requires none — which then lands on retain and consolidation as well, silently degrading their generation quality (the reported case: retain emitting facts in Chinese from an English transcript).

Closes #2998.

Change

Adds REASONING_EFFORT to the existing per-operation set, following the established fallback pattern used by every neighbouring field:

HINDSIGHT_API_RETAIN_LLM_REASONING_EFFORT         -> falls back to HINDSIGHT_API_LLM_REASONING_EFFORT
HINDSIGHT_API_REFLECT_LLM_REASONING_EFFORT        -> falls back to HINDSIGHT_API_LLM_REASONING_EFFORT
HINDSIGHT_API_CONSOLIDATION_LLM_REASONING_EFFORT  -> falls back to HINDSIGHT_API_LLM_REASONING_EFFORT

An operator can now set reflect to none (satisfying the tool-call constraint) while retain keeps low.

• config.py: three ENV_* constants, three str | None dataclass fields, three from_env() parses (empty/unset → None).
• memory_engine.py: the three per-operation LLM configs read config.<op>_llm_reasoning_effort or config.llm_reasoning_effort. The global default config is unchanged.
• docs: new rows in configuration.md (and its skill mirror).

Tests

TestPerOperationReasoningEffort in test_per_operation_llm_config.py:

• per-operation overrides are parsed from env, global left untouched;
• unset per-op fields stay None (fall back at runtime);
• MemoryEngine threads each override into its LLM config, and an unset operation inherits the global value.

Notes

• The per-operation override env vars are intentionally not enumerated in .env.example — matching the existing precedent (none of the ~30 other per-op override vars are listed there; the file documents the prefix-fallback pattern once and lists only the global HINDSIGHT_API_LLM_REASONING_EFFORT).