fix(llm): forward consolidation temperature

[lusoris]

Summary

Forward llm_temperature_consolidation to both observation-consolidation LLM calls:

• the batch synthesis call (scope="consolidation")
• the semantic dedup adjudication call (scope="consolidation_dedup")

The configuration field and its environment resolution were added by #2469 for #2459. The current call sites do not pass the resolved value, so HINDSIGHT_API_LLM_TEMPERATURE_CONSOLIDATION has no effect on these two operations. This also means the global none setting cannot omit temperature consistently across all consolidation calls.

Changes

• Pass the resolved consolidation temperature to both LLM call sites.
• Add a focused batch-call regression test with a non-default temperature.
• Assert that dedup adjudication receives the configured default.
• Complete the minimal test config objects used by existing consolidation tests.

Verification

• pytest tests/test_consolidation_retry_budget.py tests/test_consolidation_dedup.py -q — 52 passed.
• pytest tests/test_consolidation.py::TestBuildResponseModel -q — 8 passed.
• pytest tests/test_llm_temperature_pipeline.py -q — 3 passed.
• ./scripts/hooks/lint.sh — passed.
• Repository pre-commit hooks — passed. The existing vulture and knip findings remained advisory and do not involve this diff.

Risk

Low. This uses the existing resolved configuration and provider behavior. The default remains 0.0. A configured None continues through LLMConfig.call() so providers can omit the parameter for models that reject explicit temperatures.

Related

• Refs Fix LiteLLM provider temperature handling for GPT-5.5 #2459
• Completes the consolidation wiring introduced in fix(llm): make per-operation temperature configurable (#2459) #2469

AI assistance

OpenAI Codex with the gpt-5 model, high reasoning, tool use, and code execution. The client does not expose the context-window size.

[koriyoshi2041]

Checked exact head 467e0987 locally. The consolidation wiring and focused regressions behave as intended: 60 tests passed across retry-budget, dedup, temperature-pipeline, and response-model coverage.

The remaining 3 temperature-pipeline cases stopped during fixture setup before exercising this patch because the local environment points at a SOCKS proxy without socksio; this is an environment dependency error, not an assertion failure. In particular, the new 0.65 batch assertion and the dedup temperature assertion both passed.