[FIX] Preserve configured LLM in non-batch fallback extractors

[gillespied]

Description

_run_non_batch_extractor in BatchLLMPropositionExtractorSync and BatchTopicExtractorSync builds its fallback extractor without passing the configured llm, so it silently defaults to GraphRAGConfig.extraction_llm (us.anthropic.claude-sonnet-4-6) instead of whatever LLM was explicitly configured.

Fixes #534.

Changes

• batch_llm_proposition_extractor_sync.py: pass llm=self.llm to the fallback LLMPropositionExtractor
• batch_topic_extractor_sync.py: pass llm=self.llm to the fallback TopicExtractor
• Regression tests for both classes asserting the fallback extractor is constructed with the extractor's configured llm

Problem

Related issue: #534

When a node set falls below Bedrock's minimum batch size (100 records), both batch extractors fall back to a non-batch extractor. That fallback ignores self.llm entirely, so it uses GraphRAGConfig.extraction_llm (DEFAULT_EXTRACTION_MODEL = 'us.anthropic.claude-sonnet-4-6') regardless of what LLM was configured. Outside a region where that us.* profile resolves, this raises ValidationException: The provided model identifier is invalid. We hit this in eu-west-2 with an explicitly configured eu.anthropic.claude-sonnet-4-6 LLM.

This is a different mechanism from #344/#347 (pickled BedrockConverse losing region_name in a spawned worker) — here the configured LLM is never passed to the fallback at all, so the global.* inference profile workaround for #344 doesn't apply.

Testing

• Unit tests added/updated — one regression test per class, patching the target extractor class and asserting it's constructed with llm=<the configured llm>
• Integration tests added (as appropriate) — not applicable, this only affects extractor construction, no AWS call needed to reproduce
• Existing tests pass (pytest) — ran tests/unit/indexing/extract/test_batch_llm_proposition_extractor_sync.py and test_batch_topic_extractor_sync.py (11 passed), and the full tests/unit suite (2137 passed; one pre-existing, unrelated error in test_integ_dependency_compatibility.py caused by no pip binary in my venv, reproduced identically on main without this change)
• Tested manually (describe below)

Manually reproduced against a real eu-west-2 account: a BatchConfig-configured eu.anthropic.claude-sonnet-4-6 extraction, where one ProcessPoolExecutor worker's chunk count landed under 100 and hit this exact fallback path, raising the ValidationException quoted in #534.

Checklist

• Code follows existing style and conventions
• License headers present on new files (no new files; existing headers preserved)
• Documentation updated (if applicable) — not applicable, internal fallback behavior only
• No breaking changes (or clearly documented) — additive llm= kwarg passthrough only

[noel-improv]

@gillespied thanks for your contribution so far these looks good. Both extractors take LLMCacheType, so self.llm is the right thing to pass here.

[mykola-pereyma]

Thank you for the contribution! Clean, well-scoped fix — passing llm=self.llm into both non-batch fallbacks correctly preserves the configured inference profile, and the per-class regression tests pin the behavior nicely. Verified these are the only two fallback sinks (no non-sync siblings share the omission). LGTM.

[oussamahansal]

LGTM