fix(machine): decode QR at intrinsic frame + tuned capture resolution

The camera pairing source decoded frames at the <video> element's CSS box
size (qr's readFrame default) rather than the intrinsic frame, and let the
stream stay at the panel-bound ~720p that frontalCamera negotiates from the
screen size. On the 1280x800 kiosk with a fixed-focus 5MP scan camera that
left far too few pixels-per-module for a dense spire-seed QR, so a centered,
in-square code never decoded.

Decode the intrinsic frame (readFrame fullSize=true) and pin a deliberate
1280x960 capture via applyConstraints. lamassu-machine caps QR scanning at
640x480 for decode speed (megapixels only slow the per-frame decode); our
seed QR is denser than a lightning invoice, so 1280x960 balances
pixels-per-module against latency and keeps auto-exposure from blowing out a
frame-filling phone screen.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

apps/machine/src/services/pairing/qr-source.ts

@@ -32,6 +32,31 @@ export class QrPairingSource implements PairingSource {
     if (!video) throw new Error('QrPairingSource requires a <video> element')
 
     const camera = await frontalCamera(video)
+
+    // `frontalCamera` requests `ideal: screen.{width,height}`, so on the kiosk's
+    // 1280x800 panel it otherwise streams at ~720p — and `decodeQR` then
+    // center-crops to a square, leaving too few pixels-per-module for a dense
+    // spire-seed QR on this fixed-focus lens. Pin a deliberate 1280x960 capture
+    // instead: lamassu-machine caps QR scanning at 640x480 for decode speed
+    // (megapixels just slow the per-frame decode), but our seed QR is denser
+    // than a lightning invoice, so 1280x960 is the balance — ~14-18px/module at
+    // frame-fill, still fast, and it meters exposure better than maxing the
+    // sensor (a frame-filling QR keeps auto-exposure from blowing out on a
+    // bright phone screen). The stream lives on the <video>'s srcObject
+    // (QRCamera.stream is private); soft `ideal` so a camera that can't honor
+    // it degrades to its closest mode instead of throwing.
+    try {
+      const stream = video.srcObject
+      if (stream instanceof MediaStream) {
+        await stream.getVideoTracks()[0]?.applyConstraints({
+          width: { ideal: 1280 },
+          height: { ideal: 960 },
+        })
+      }
+    } catch (e) {
+      onError?.(e)
+    }
+
     const canvas = new QRCanvas() // decode-only; no overlay canvases needed
 
     let stopped = false
@@ -46,7 +71,11 @@ export class QrPairingSource implements PairingSource {
     cancel = frameLoop(() => {
       if (stopped) return
       try {
-        const result = camera.readFrame(canvas)
+        // `fullSize: true` decodes the camera's intrinsic frame (videoWidth ×
+        // videoHeight) rather than the <video> element's CSS box — the default
+        // (`false`) was decoding the few-hundred-px on-screen preview, which
+        // (compounded by `object-cover` cropping) starved the decoder.
+        const result = camera.readFrame(canvas, true)
         if (result) {
           stop()
           onScan(result)