feat: read fingerspelled LSE letters from the camera

[Endika]

The camera now works and the app spells. First engine end to end: MediaPipe → segmenter → handshape table → stabiliser → transcript.

What landed

• MediaPipeLandmarkSource — camera → 21 hand landmarks, WASM served same-origin. Releases the tracks on stop so the camera indicator actually goes off.
• HandshapeAlphabetClassifier — scores the live handshape against a table of LSE fingerspelling shapes. No dataset, no download, useful on first launch.
• RecognizeSignsUseCase — routes frame engines (a held letter) and window engines (a travelled sign) differently, each with its own stabiliser.
• Container — constructor injection, no framework.

Two things worth a look

The fixture was lying. It modelled a curled finger as a straight finger pointing backwards, so the joint angle stayed at 180° and every handshape measured as fully straight — curl was 0.00 for every input. Probing before writing assertions is the only reason this was caught; tests written against it would have passed while measuring nothing.

Averaging alone picks confident wrong letters. d and l differ only in the thumb, so a completely wrong thumb cost one fifth of the score and l won on its three matching curled fingers. Any feature outside tolerance now vetoes the letter outright, so an ambiguous shape returns nothing. A missing letter is obvious to the signer; a wrong one is not.

Deliberately not recognised

g h j m n ñ p q r t x z — traced through the air, or dependent on palm orientation, which the invariant features discard on purpose. Listed in the UI so the gap is visible rather than mysterious. They are the trained model's job.

Camera untested under WSL2 (no device access); the pipeline's logic is covered by 43 tests against synthetic hands.