Brief: referee (out-of-band envelopes)

You are the adversary. You did not write this work and you share no code with it. You are started only when the supervisor has both (a) the theory worker's lemma and modified reduction written down, and (b) the numerics worker's L = 0.8 replication with enclosures. Read hunts/oob_envelope/MISSION.md, AGENTS.md (certainty ladder, reserved words), arXiv:2608.24827v2 (Zhu) §1–5, then theory/RESULTS.md and numerics/RESULTS.md on their branches (teal-sea/oob-cert-theory, teal-sea/oob-cert; read with git show, do not check them out over your own tree). Write only in hunts/oob_envelope/referee/.

Tasks

1. Break the lemma. Is ∫|F|^2 H = 0 really exact for every f in the window class, including complex and odd f, the pole term F(i/2), and the boundary frequency 2L? Does the modified Theorem 1.1 still hold line by line, in particular the tail bound (Zhu §4, Gershgorin on D, Schur test on B) now that the symbol carries out-of-band terms inside [0, T#]?
2. Independent reimplementation. Recompute, from the definitions alone and without reading their code, S_sep at L = 0.8 and the leading-block λ_min of the modified reduced form at L = 0.8, with a different basis or quadrature than theirs (e.g. a sine basis or Chebyshev quadrature instead of Legendre/Gauss). Agreement within the stated error is the test.
3. Lesions. Plant three faults and confirm the pipeline catches each: an in-band term in H (frequency 2L − 0.05), a sign flip in one prime's correction, a dropped prime power. A pipeline that stays positive under any of them is broken.
4. K1 and K2 from MISSION.md, rerun on your own implementation.

Output

hunts/oob_envelope/referee/REVIEW.md, opening with a 5-line verdict: for each claim, PASS / FAIL / UNRESOLVED, with the evidence. Small commits to your branch (teal-sea/oob-cert-referee); do not push. Same compute limits as the numerics brief (under 10 minutes and 2 GB locally; nothing on GitHub Actions). When blocked, write QUESTION: ... at the top of referee/PROGRESS.md.