Answer-quality fixes (review, area 3)

[adamjohnwright]

These come from the max-level review of the agent and retrieval code. Each one was reproduced by the reviewer.

Finding	Fix	Measured or tested
Disease-variant keyword search indexed all 21 columns, while the vectors embedded 12. RRF never fused the two, and results came back doubled.	EMBEDDED_CONTENT_COLUMNS: keyword search indexes the embedded columns, limited to those present in the file	On the real Release 97 bundle, 300 of 300 vector documents now match the keyword text (0 of 300 before)
Disease-variant citations had empty labels	Labelled with the variant metadata	Unit test
A failed grader or web search failed a turn that had already answered	Logged and dropped	Unit test
A safety-check refusal raised an error instead of giving the polite refusal	OpenAIRefusalError is treated as an unsafe verdict	Unit test
User-guide answers ignored the reader's language	The user-guide prompt gets LANGUAGE_INSTRUCTION, in the same place as the Reactome prompt	3 runs each, old vs new prompt: French 0/3 → 3/3; Spanish 3/3 → 3/3; English 3/3 → 3/3. Answer sweep 16/16.
The classifier test compared the prompt with itself, so it could never fail	Replaced: the no-MCP prompt must not mention live services	Sabotaged: always including the live rule fails it

./checks.sh passes.

🤖 Generated with Claude Code