feat(config): per-operation llm_extra_body overrides

[nicoloboschi]

Problem

extra_body could only be set globally, via HINDSIGHT_API_LLM_EXTRA_BODY, so retain, reflect and consolidation were forced to share one dict — unlike every other LLM param, which already has a per-operation override.

That's limiting for knobs that are genuinely per-operation. The motivating case: disabling a model's thinking mode for retain extraction while leaving it on for reflect (vLLM chat_template_kwargs). Same for a different max_tokens or top_p per operation.

Change

Three new env vars, following the reasoning_effort pattern exactly:

Variable	Falls back to
HINDSIGHT_API_RETAIN_LLM_EXTRA_BODY	HINDSIGHT_API_LLM_EXTRA_BODY
HINDSIGHT_API_REFLECT_LLM_EXTRA_BODY	HINDSIGHT_API_LLM_EXTRA_BODY
HINDSIGHT_API_CONSOLIDATION_LLM_EXTRA_BODY	HINDSIGHT_API_LLM_EXTRA_BODY

Each is parsed into an optional HindsightConfig field and resolved in MemoryEngine as config.<op>_llm_extra_body or config.llm_extra_body, so an operation left unset keeps using the global value. Static server-level config (not added to _CONFIGURABLE_FIELDS), matching the global flag.

# thinking off for extraction, on everywhere else
HINDSIGHT_API_LLM_EXTRA_BODY='{"chat_template_kwargs": {"enable_thinking": true}}'
HINDSIGHT_API_RETAIN_LLM_EXTRA_BODY='{"chat_template_kwargs": {"enable_thinking": false}}'

A per-operation value replaces the global dict rather than merging with it. Extra-body params are provider-native, so a partial override merged onto a global dict aimed at a different provider would mix two providers' field names — and replacement is how every other per-operation override already behaves.

Scope notes

• No behaviour change when the new vars are unset: each per-op config resolves to config.llm_extra_body, exactly as before.
• Indexed multi-LLM members (..._LLM_<n>_EXTRA_BODY) already have their own per-member extra_body and are untouched.
• No API/schema surface, so no OpenAPI or client SDK regeneration.

Tests

New TestPerOperationExtraBody in tests/test_per_operation_llm_config.py, mirroring TestPerOperationReasoningEffort one-for-one:

• test_per_operation_extra_body_from_env — each override parses, global untouched
• test_per_operation_extra_body_fallback_to_global — unset stays None
• test_memory_engine_applies_per_operation_extra_body — engine threads the reflect override through and retain/consolidation fall back to the global value

Verified locally: 78 passed across test_per_operation_llm_config.py + test_llm_extra_body.py + test_multi_llm_config.py; ./scripts/hooks/lint.sh and uv run ty check hindsight_api/ clean.

Docs

Three rows added to the per-operation table in hindsight-docs/docs/developer/configuration.md.