verification: two Codex findings on #13 at 371c512 deferred from the final turn #153

Two findings from Codex's review of #13 at 371c512, listed here and not fixed, as the manager decided for the final #13 turn.

1. Non-check assumptions are never compared with the rules (4114676179, verification/hsverify/resultcheck.py, assumption_mismatches). The ETH3D case declares corner_allowance and battery_width because its expected sweep ranges depend on them, but the check only matches check ids and their threshold_ft. If either rule changes, the mismatch isn't reported as an assumption change, so the case's expectations can fail or pass for the wrong reason.
2. Portrait replays pass replaycheck (4114676183, verification/hsverify/replaycheck.py). When a replay's JPEG and its declared w/h are both portrait, the size check passes and the intrinsics can fit, although downstream projection assumes the landscape sensor image (AGENTS.md: rotate the image and you must rotate the intrinsics). Reject portrait keyframes, or require the intrinsics to match the rotation.

Other deferred verifier items are in #125.

🤖 Generated with Claude Code