feat(engine): add per-claim confidence persistence (#465)

[manavgup]

Summary

• Update the compiler prompt to request subjects and source_ids per claim, enabling per-claim source attribution from LLM output
• Pass source UUID in user prompt metadata so the LLM can reference it in claim source_ids
• Add integration tests verifying claim-level confidence aggregates correctly into article-level confidence

Context

PR #800 shipped Phase 1 of issue #465 (article-level confidence with persistence). This PR advances the issue by closing the gap between the LLM prompt and the existing CompiledClaim table schema — the table already had subjects, source_ids, and confidence_score fields, but the compiler prompt never asked the LLM to populate them. Now it does.

What was already in place (from PR #800 and #784):

• CompiledClaim table with confidence_score, source_ids, last_reinforced_at, subjects
• compute_claim_confidence() and aggregate_claim_confidence() pure functions
• _persist_claims() storing claims with per-claim confidence scores
• _refresh_confidence_score() aggregating claim scores to article level
• GET /wiki/articles/{id}/claims endpoint and service layer
• Full DTOs, migrations, and unit tests

What this PR adds:

• Compiler prompt schema now includes subjects and source_ids fields per claim
• Prompt rules instruct the LLM to set subjects (1-3 canonical entity names) and source_ids (UUIDs from metadata)
• _build_user_prompt passes Source ID: <uuid> in metadata block
• Integration test: article confidence is the mean of its claim scores
• Integration test: _persist_claims correctly stores subjects and source_ids from DTOs
• Unit test: _build_user_prompt includes source ID in output

Not included (deferred to future PRs):

• Concept identity / embedding clustering
• Cross-article claim deduplication

Test plan

• All existing unit tests pass (35 compiler tests, 24 claim confidence tests, 30 citation tests)
• All existing integration tests pass (3 confidence integration tests)
• New unit test: test_build_user_prompt_includes_source_id verifies source ID in prompt
• New integration test: test_article_confidence_aggregated_from_claims verifies claim-to-article aggregation
• New integration test: test_persist_claims_stores_source_ids_and_subjects verifies claim persistence
• ruff check passes with no errors
• mypy passes (4 pre-existing errors in unrelated files)

🤖 Generated with Claude Code