Leveson concern: no end-to-end test for failed-extraction propagation through synthesis

[melek]

Problem

When EXTRACT_FIELD oracle contract validation fails, recovery sets value="extraction_failed" and confidence="low". This record propagates to Phase 5 synthesis. The SYNTHESIZE_CLAIM contract checks data_completeness and requires qualification for low-completeness papers — but there is no test verifying this end-to-end chain.

A plausible failure mode: a paper with multiple failed extractions gets cited in the review without qualification, making claims appear better-supported than the evidence warrants.

Proposed Fix

Add an integration test that:

1. Creates a mock workspace with extraction records containing extraction_failed values
2. Runs validate_synthesis_claims() against a review that cites those papers without qualification
3. Verifies the validator catches the missing qualification
4. Verifies the validator passes when qualification is present

Also: when Phase 4→3 feedback loop terminates at max_feedback_iterations with conceptual saturation still above θ_c, the progress line should include a warning — not just the neutral "returning to Phase 3" message. The current reporting understates the risk of an incomplete concept space.

Context

Leveson (STAMP/STPA Safety) panelist concern during panel review (2026-03-23). The individual pieces work (oracle contract recovery is tested, synthesis claim validation is tested) — the gap is the composition.