Defect 2 (residual): pro-cyclical APPROVE filter hides thin output

[Jodre11]

Context

Part of the "reproducible borderline reviews" flow that began with a real-PR verdict-flip incident (the same pipeline returned APPROVE on one run and CHANGES_REQUESTED on another, on identical input). Post-incident analysis localised three layered defects. Defect 1 (single-draw specialist recall variance) shipped in #60 (variance-resampling on a boundary gate). This issue tracks the residual of Defect 2.

The defect

Pro-cyclical APPROVE filter. Under an APPROVE verdict the orchestrator suppresses sub-75-confidence findings from the posted review (isPosted(), posting policy in verdict-rubric.md). A low-recall run is more likely to land on APPROVE — so the engine hides its own thin output precisely when that output is thinnest. The filter's noise-suppression value is real on genuinely clean PRs; the problem is it can't distinguish "clean" from "thin."

What #60 already did

• The boundary gate absorbs much of the amplifier: a thin APPROVE with borderline findings now triggers a 2nd stochastic draw instead of shipping silently.
• The sub-75 disclosure line (N finding(s) below the posting threshold — see synthesiser report.) addresses the honesty half — an APPROVE no longer looks cleaner than the run was.

What's left (this issue)

Decide whether the residual pro-cyclical dynamic needs a structural change beyond the gate + disclosure, or whether it's now adequately mitigated. This is best judged from organic observation of real reviews (see watch items below), not a speculative change now.

Watch items that would reopen active work here:

• Does the gate stay quiet on genuinely clean-cut PRs (cost containment)?
• Does an APPROVE with a disclosure line still feel like it's hiding substance, in practice?
• Does B2's bare contested-presence trigger over-fire on low-value contested findings? (If so, add a confidence floor to the B2 predicate in review-core.mjs and update test_gate_fires_b2_contested.)

Status

Low urgency. Mitigated by #60; open pending organic signal. Do not change speculatively.