revert(search): restore baseline answer prompt (commit-policy regressed)

[us]

Benchmarked #71's commit-policy prompt on the 50-q SimpleQA+FRAMES sample: NET REGRESSION — overall 64%→48%, INCORRECT 2→17 (forcing commits caused confident-wrong on multi-hop + weak retrieval, destroying crw's near-zero-hallucination property). Revert to baseline. Real lever is retrieval recall, not prompt-forced commits.

[Copilot]

Pull request overview

Reverts the /v1/search answer synthesis system prompt in crw-extract back to the pre-#71 “baseline” wording after benchmarking showed the commit-policy prompt increased confident-wrong answers and reduced overall accuracy.

Changes:

• Removes the “COMMIT to a specific answer” commit-policy instructions from the answer system prompt.
• Restores baseline guidance: abstain when sources don’t cover the query and write a neutral prose answer.

Comments suppressed due to low confidence (1)

crates/crw-extract/src/answer.rs:37

• The prompt rules still instruct the model to call a cite_sources tool, but this implementation actually requires the model to append an ===CITATIONS=== JSON block (see the augmented_prompt instructions and parse_answer_and_citations splitting on ===CITATIONS===). Keeping tool-call wording here creates contradictory instructions within the same system prompt and can increase malformed outputs / citation parse failures. Consider updating this section to describe the ===CITATIONS=== block instead of a tool call.

- Write a direct, neutral answer in 3–6 sentences of plain prose.
- After producing the answer, you MUST call the `cite_sources` tool to
  report which sources you used. Each citation gives a `source_id` (the
  integer index of the source) and a `position` (a hint for ordering;
  use the position the source had in the input list).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.