Health: colour the swap bar by paging rate, not by fill

[cyberb]

Problem

A device that has been up for months parks cold pages in swap and never reclaims them. Reported on the forum as a Nextcloud problem (nextcloud very slow, 100% cpu usage), but reproduced on my own box where nothing is actually wrong:

Mem:  7864 total, 5390 used, 2210 buff/cache, 1597 available
Swap: 3000 total, 3000 used, 0 free

procs -----------memory---------- ---swap-- ...cpu-----
 r  b   swpd   free   buff  cache   si   so  ... us sy id wa
 0  0 3073016 237732 240144 2054684   0    0  ...  5  3 93  0

si/so are 0 across the sample, CPU 93% idle, 1.6 GB available. But pctStatus() turned the swap bar red at >=90% fill, so the Health page showed a permanent alarm — right next to its own "swap in 0 · swap out 0 KB/s" line.

Change

Swap fill says how much has ever been paged out; it says nothing about whether the machine is struggling now. The rate does, and backend/health/metrics.go already collects swap_in_pages/swap_out_pages.

• The swap bar's status now comes from swapRate — 256 KB/s warns, 2 MB/s is an exception.
• Fill stays as the bar's length and in the used / total MB readout, where it informs without rendering a verdict.
• Memory and disk bars are untouched; fill is the right signal there.
• The dev stub never emitted the two counters, so the rate line read 0 in dev regardless. It now advances them.

Verified on a real device

Deployed to a box running 19 apps (platform rev x1, all 75 snap services active). Swap 3000/3000 MB, pswpin/pswpout identical across a 2s sample, so the bar renders green at 100% fill with 0 · 0 KB/s beneath it. Before this it was red.

Also included: visual_diff_skip_build 3076 → 3084

Not related to the swap change, but this branch cannot go green without it. The nightly artifact prune on the CI host deleted platform/3084-*, which is the build FindLatestBuild("stable") resolves to, so ci-diff downloads an empty tree and fails on every branch build:

Stable build: #3084
  SKIP: base=missing compare=artifact/playwright/desktop/screenshot
FAIL: no screenshots were compared stable

3076 was the previous occurrence. The constant only takes effect when it equals the current latest stable, so it went inert when stable moved to 3084.

Note this makes visual-diff pass by returning early rather than by comparing. No e2e spec visits /health, so this PR loses no coverage it ever had — but the underlying defect remains: visual-diff/cmd/app.go:54 returns nil for a build directory that does not exist, making a pruned baseline indistinguishable from a screenshot regression. Worth a follow-up.

The baseline regenerates on the next stable push and is protected from pruning from then on by the artifact cleaner now running on the CI host, which keeps the last green stable per app regardless of age.

Tests

New tests/unit/Health.spec.js covers full-but-idle (green), sustained paging with swap mostly free (red), light paging (amber), and that pctStatus still governs the other bars. Full suite: 25 suites / 129 tests green, eslint clean. Jsonnet validated with drone jsonnet + drone lint.