refactor(sleep): one stage-key canonicalisation, not two kept in parity by comment

[ryanbr]

Follow-up to #733, found reviewing it.

Two implementations of one rule, and the code already knew. stageColorFor did its own trim().lowercase() and folded "awake", "wake" in a when branch, while canonicalStage() in SleepStageTimelineLogic.kt does exactly the same — carrying the comment "(stageColorFor parity)". A drift risk documented instead of removed.

// before
private fun stageColorFor(name: String): Color = when (name.trim().lowercase()) {
    …
    "awake", "wake" -> Palette.sleepAwake

// after
private fun stageColorFor(name: String): Color = when (canonicalStage(name)) {
    …
    "awake" -> Palette.sleepAwake

Both files are package com.noop.ui and canonicalStage is internal, so no plumbing. Adding a stage alias is now a one-place edit — before, a new alias in canonicalStage would silently fail to pick up a colour, which is the quiet-wrong-answer failure mode rather than a visible one.

Behaviour-identical, verified across every input the old branch accepted: "deep", " REM ", "Light", "awake", "WAKE", "wake", whitespace variants, empty string and unknown all resolve to the same tint before and after, fallback included.

#733 didn't create this — it made it visible by moving the two halves into adjacent files, which is a good argument for that refactor rather than against it.

Android only, no schema, no scores, no strings; i18n clean. Not compiled — android.yml is disabled — though this is a two-line change to a when subject within one package.

[ryanbr]

Re-review found two things in my own change, both now pushed.

1. A KDoc link that could not resolve. [stageColorFor] from SleepStageTimelineLogic.kt points at a private function in another file — not in scope there, so Dokka would warn. The reverse link is fine: [canonicalStage] is internal in the same package. The asymmetry is easy to miss because the two links look identical. Now a code span plus the filename.

2. The last (stageColorFor parity) note, in StageRowSpansTest.kt:33. That phrase described a hand-maintained invariant, which is exactly what this PR removes — leaving it means grepping it turns up a stale hint. The test now covers the colour path as a side effect of covering the spans, so it says that instead.

No behaviour change in either. CI green, three commits.

[ryanbr]

Third pass, one more of my own: I had added a doc block above the existing one on stageColorFor instead of editing it, leaving two consecutive /** */ on the same declaration. Kotlin takes only the last as KDoc, so the original case-insensitive note silently became a dangling comment. Duplicate documentation, in a PR about removing duplicate logic.

Merged into one block, keeping the original wording and adding the unknown-stage fallback, which neither block stated.

CI green, MERGEABLE/CLEAN, +13/-5.

Out of scope here, but the same defect exists in two other places and I put both there: PuffinExperiment.kt:118 (harmless — near-identical text) and HrvAnalyzer.kt:342, where the detached block is the one reading "Byte-parity twin of Swift classifyCoverage". That is a parity-contract note that is no longer attached to the function it documents. Happy to fix both in a separate one-liner PR.