iOS: <Camera> records 4K — preview resolutionBias overrides the video output's targetResolution (failing repro)

[radko93]

Relates to #3963. Builds on #3965.

#3965 is a great root fix for the imperative case in #3963 — switching from a summed weighted penalty to a lexicographic penalty vector means a higher-priority constraint can't be numerically swamped by a lower-priority one. 👍

But it makes constraint order decisive, and that exposes a second path #3965 doesn't cover: the auto-injected preview resolutionBias. This PR adds a failing harness test for it (test-only — the Device Farm run is the reproduction). A proposed fix is in the <details> at the bottom.

The gap

In real <Camera> usage the resolver receives constraints in roughly this order (Camera.tsx prepends the preview to outputs; useCameraController.ts appends one { resolutionBias } per output):

[ ...userConstraints, { resolutionBias: previewOutput }, { resolutionBias: videoOutput }, ...internal ]

So the preview bias outranks the video bias. HybridCameraPreviewOutput.targetResolution is .min(screenSize) with streamType = .video, which gives every ≥screen-size format (4K) penalty 0 and penalizes the requested 720p/1080p. Under #3965's lexicographic comparison, for any formats that tie on fps the preview term is compared before the video term and still selects 4K — targetResolution is overridden again.

#3965's harness test uses [{ resolutionBias: videoOutput }, { fps: 60 }] with no preview, so this path is currently untested.

Test

Adds an it(...) to visioncamera.constraints.harness.ts that mirrors <Camera>: a preview output + a Frame output (targetResolution 720p), with [{ resolutionBias: preview }, { resolutionBias: frame }]. It asserts the delivered Frame is not ~4K. Frame dimensions are the only device-readable proof of the negotiated format (CameraSessionConfig exposes no resolution). Expected to fail on main and also with #3965 applied — happy to rebase onto #3965's branch to show that directly. Repro device: iPhone 17 Pro / iOS 26.5.1, front camera (camera-agnostic — any device with a format larger than the screen).

Proposed fix (preview should not outrank capture outputs) — direction for discussion, not verified

The .min(screenSize) rule is correct for a preview-only session (you want a screen-sized format there), so the fix is about priority, not the rule: when a capture output is present, its resolution target should outrank the preview's.

Under #3965's lexicographic model that just means ordering the auto-appended biases so capture outputs come before preview outputs. In useCameraController.ts:

const stableConstraints = useMemo<Constraint[]>(() => {
  const captureBiases = stableOutputs
    .filter((o) => !isPreviewOutput(o))
    .map<Constraint>((o) => ({ resolutionBias: o }))
  const previewBiases = stableOutputs
    .filter(isPreviewOutput)
    .map<Constraint>((o) => ({ resolutionBias: o }))
  // capture targets first (higher priority), preview last
  return [...constraints, ...captureBiases, ...previewBiases]
}, [/* ... */])

(Equivalently, drop the preview's resolutionBias entirely when a capture output is present — a preview can downscale from any format.) I haven't verified the exact JS predicate/ordering on a device, so treat this as the direction rather than a finished patch.

[vercel]

@radko93 is attempting to deploy a commit to the Margelo Team on Vercel.

A member of the Team first needs to authorize it.