The plate is authored y-up while canvas text is laid out y-down, so the affine
solved from the raw local corners had a negative determinant: every verdict drew
mirrored, with the line order reversed, in idle and reveal alike.
Solve the transform in a y-down text space instead, flip the clip to match, and
cover the orientation with a regression test.
A failed upstream call returned HTTP 200 with a hash-derived answer that the UI
rendered with the same P(FAVORABLE) and CONFIDENCE figures as a real decision.
The fallback is now badged OFFLINE with those derived numbers blanked out.
Request errors also rendered "SIGNS POINT TO YES" and always reported a network
timeout. Failures now show CANNOT PREDICT and the real HTTP status or backend
message, and the stored history is validated before use so a malformed value
cannot break the whole interface.
