feat(examples): auto-sweep mode for stress-tv (find the FPS sweet spot)

[chiefcll]

What

Adds an automatic sweep to the stress-tv example. Instead of driving the grid by hand with the remote, ?autosweep=true ramps card count across every tier, measures FPS at each step, and reports the highest count that still holds the target frame rate — the "sweet spot."

?test=stress-tv&autosweep=true              # target 60 fps (default)
?test=stress-tv&autosweep=true&targetfps=30

How it works

• Sweeps each tier (rect → +image → +title → full card) up the count ladder.
• Measures median FPS per step via rAF timing, discarding a short warm-up so the rebuild spike / first-frame upload don't skew it.
• Bisects between the last good rung and the first bad one for a sharper number.
• Reports two distinct ceilings so they're never conflated:
  • performance wall — the measured FPS knee (CPU- or GPU-bound)
  • correctness wall — the Uint16 index-buffer cap (16384 quads / glyphs)
  • reported sweet spot = min(perf, correctness); counts are clamped to the cap so the sweep never builds a scene whose own text would drop out.
• Output to console.table and an on-screen panel.

VAO A/B

VAO is fixed at renderer construction, so the runner measures the current setting and you compare two runs:

?test=stress-tv&autosweep=true   vs   ?test=stress-tv&autosweep=true&novao=true

This is the honest way to settle "is VAO on actually faster on this device" — a strict win when CPU/driver-bound, neutral when GPU/fill-bound, and occasionally a wash on flaky embedded WebGL1 VAO drivers (which is exactly what disableVertexArrayObject is for).

Verified

On a 120 Hz dev machine (target 60 fps), the sweep reported:

tier	sweet spot	limiter
1 rect	16 384	index cap
2 +img	8 192	index cap
3 +title	2 010	index cap
4 full	804	index cap

These match the analytical caps exactly. This machine is fast enough that FPS never dipped below 60, so the index cap binds on every tier; on a slower 60 Hz TV the FPS wall would bind first at lower counts, which the tool distinguishes.

Notes for reviewer

• Refactors buildGrid to take an explicit count (so the sweep can drive arbitrary values); interactive remote controls are unchanged.
• Depends on the index-buffer fix in fix(webgl): fill entire quad index buffer; cap at Uint16 limit #79. The correctness-wall math assumes the post-fix cap of 16384; without fix(webgl): fill entire quad index buffer; cap at Uint16 limit #79 the real cap is ~4167 and high-count tiers would drop text mid-sweep. Best merged after / alongside fix(webgl): fill entire quad index buffer; cap at Uint16 limit #79.

🤖 Generated with Claude Code