The palette holder has no "still valid" state

Extraction is asynchronous, and the state holder that wraps it has three states — Loading, Success, Error — with palette readable only in Success. The order inside generate() is what makes this sharp: it assigns Loading before it suspends. So for the whole duration of every generation the palette reads null, and null resolves to whatever your fallback is — usually black.

That is fine when a generation always finishes. It stops being fine the moment one can be cancelled, because a cancelled generation never assigns Success and nothing re-runs it. The holder sits on Loading, palette stays null, and the surface stays on its fallback for as long as the screen lives.

Traps

A restarting effect is a cancelling effect, and the key decides what counts as a reason. Keyed on the bitmap, a generation is only ever interrupted by a newer bitmap — which is exactly when interrupting is correct. Keyed on anything coarser, unrelated changes cancel it:

• a state container whose other fields (an extra media payload, lyrics, metadata) land after the artwork does — deterministic, because the later field always arrives later;
• a URL that goes null and comes back to the same value across a reload — the key changed twice, so the effect restarts twice, even though nothing about the image moved.

Both look like harmless extra keys. Neither adds correctness, and both add a cancellation window.

The most robust form is an effect that never restarts. Key on nothing, and let a snapshot flow decide what a new bitmap is:

// adapted — generation lives in its own effect that cannot be restarted from outside
LaunchedEffect(Unit) {
    snapshotFlow { state.bitmap }
        .filterNotNull()
        .distinctUntilChanged()
        .collectLatest { paletteState.generate(it) }
}

distinctUntilChanged drops the same-bitmap re-emissions that a recomposition produces, and collectLatest cancels only for a genuinely different bitmap. Nothing else in the composition can reach it.

A guard assigned after a suspension is not a record of "done". The common shape is generate(bm); generatedFor = url inside an if (generatedFor != url). On the cancelled path the assignment never runs, so the flag and the holder now disagree about what happened — and which way that fails depends on polarity and timing, which is the worst property a guard can have. If a flag is genuinely needed, set it before the suspension and clear it on failure, or drop it and let distinctUntilChanged do the deduplication structurally.

Hold the last colour that actually resolved. This is the fix that makes the rest of it stop mattering, and it is one extra piece of state:

// adapted — reading paletteState.palette straight paints black for the whole of every generate()
var pageBackground by remember { mutableStateOf(Color.Black) }
LaunchedEffect(paletteState.palette) {
    paletteState.palette?.let { pageBackground = it.toBackground() }
}

Every surface reads pageBackground, never the holder. A generation in flight then changes nothing on screen, and a cancelled one costs the update, not the colour.

Cancellation is not the only way the palette never arrives. If the bitmap itself never materialises — a remote image whose load fails, so the success callback never fires — the generator is never called at all and the same fallback is shown, with a completely different cause and the same symptom. Check which one you have before fixing the effect; see image-fallback-url-retry-in-composition.

"Sometimes black" is the signature, and it is why this survives. A deterministic failure gets fixed. This one depends on which async result wins a race, so it reproduces on some items, on some navigations, on a cold start but not a warm one. Treat any intermittent artwork-derived colour as this until proven otherwise, rather than as a rendering glitch.

Everything downstream of the palette inherits the fallback. A single accent is obviously wrong when it is wrong. A whole tonal scheme derived from the seed degrades to a plausible-looking scheme in the app's default colour, which reads as a design decision — see artwork-seeded-dynamic-scheme, and artwork-palette-theming for what to do when extraction genuinely returns nothing. A screen painting itself from artwork is also a screen that must be forced dark regardless of the app theme (force-dark-immersive-subtree); the two omissions land on the same screens and look nothing alike.

Verifying it

1. List every generation and the key of the effect around it. Read the key on each line, not the call:

   grep -rn --include='*.kt' -B 6 "State.generate(" . | grep -v '/build/' \
     | grep -E "LaunchedEffect|State.generate\("
   

   Thirteen generations here. Eight are keyed on bitmap alone and one on Unit with a snapshot flow inside; every other key is a cancellation source that needs a reason.

2. Find surfaces reading the holder directly instead of a held colour:

   grep -rn --include='*.kt' -E "paletteState\.palette|\.palette\?\." . | grep -v '/build/'
   

   Hits inside a LaunchedEffect that assigns a remembered colour are the correct shape. A hit inside a modifier, a background(...) or a theme call paints the fallback during every generation.

3. Reproduce it deliberately rather than waiting for it. Add a temporary delay(2000) before generate(...), then navigate onto the screen and immediately trigger whatever else updates that effect's key — a reload, a range change, a second data field arriving. With a coarse key the surface stays on its fallback; with the bitmap-only key it resolves. Remove the delay afterwards.