fix: report stale structure baselines, and correct three of them (#227)

[BernardJen]

Adds the baseline-staleness half of #227. Found three real problems on its first run.

What it found on main

Saver	baseline	measured	ratio
Starfield Warp	0.0011	0.1423	129.4×
Mandelbrot	0.0020	0.0557	27.8×
Voronoi	0.0341	0.1557	4.6×

All three are leftovers from the rewrites in #212, #210 and #211. Each rewrite made its
saver denser; the collapse check only fires on a drop, so nothing complained.

This changes what #227 is about

I named Starfield Warp in that issue, alongside Plasma, as a saver too dim for the check
to protect — its baseline sat under STRUCTURE_MIN_ABS_DROP. That was wrong.

It isn't dim. It measures 0.1423, two orders of magnitude above the margin. Its
baseline was simply stale, left behind by the saver it replaced — thirty white dashes on
navy. Correcting it moves Starfield Warp into full protection and shrinks the known
blind spot from two savers to one: Plasma, whose 0.0001 is genuine.

A test now asserts Plasma is the only sub-margin saver, so this cannot quietly grow
again.

Reported, not failed

A stale baseline is a defect in the check's configuration, not in the frame under
test. Failing on it would turn every unrelated PR red until somebody fixed a saver they
had never touched. The exit code comes solely from the fail= count — verified by reading
scripts/shader-check.mjs, which matches only SHADERCHECK_DONE ok=(\d+) fail=(-?\d+).

The ratio is calibrated, not picked

4×. I tried 2× first and it flagged two savers whose own measurement is bimodal at
frames 5/12: Raymarch Fractal read 0.0082 in one run and 0.0185 in the next (2.3×),
Reaction Diffusion likewise 2.3×. That is exactly the effect #192 was filed over, and a
threshold has to clear the noisiest saver rather than the median.

4× still catches every real case — Truchet's historical gap was 4.1×, and the three above.

A rejected approach, measured rather than argued

#227 also asks for the flat-gradient case: a smooth ramp passes for a sub-margin saver,
because it has few edges and a wide spread — and so does a healthy Plasma frame.

I tried coarse-stride edge density. A full-frame ramp differs by stride/width between
two samples, so on a 1200-wide frame it scores zero at strides 1, 8 and 32 alike, while
a real image should not. It does separate them — but not by enough to ship:

• the lowest healthy score is DVD Logo at 0.0026 (stride 8), so any floor that catches
  a ramp sits within 2.6× of a real saver — one saver getting slightly sparser turns CI
  red;
• it only catches the slowest gradients. A steeper or repeating one scores normally, so it
  would not reliably catch the Physarum-class bug it was aimed at.

Instrumentation removed. That part of #227 stays open, with the negative result recorded
so nobody re-derives it.

Verification

• Branched from main (7c5aa56).
• npm run lint clean.
• npm test — 407 passed (400 + 7 new staleness tests).
• npm run shadercheck before the baseline fix: 3 stale reported, fail=0.
• npm run shadercheck after: stale list empty, ok=150 fail=0.
• All three corrected baselines reproduce on a second run: Starfield 0.1414 vs 0.14,
  Mandelbrot 0.0557 vs 0.055, Voronoi 0.1558 vs 0.15.
• Baselines re-sorted descending, so chore: re-sort the structure baselines and guard the order with a test #224's ordering test still holds.

Not verified

• The 4× ratio clears the two noisy savers I know about. If another saver turns out to be
  bimodal by more than 4×, it will produce a spurious warning — visible, and not CI-breaking.
• The flat-gradient case for Plasma remains open, as above.