Make wear and jitter tuning knobs honest about their ranges

[malpern]

Summary

Follow-up to #1243, addressing its review findings (the PR auto-merged on the flaky-test re-run before these landed):

• Drop the accidental saturate() on wearStrength (the 1.5 default silently behaved as 1.0; default is now an honest 1.0 — the exact effective value the approved goldens rendered)
• Gate seed-only gloss variation (roughness offset, specular gain) with min(highlightJitter, 1) so zeroing the knobs fully restores the uniform baseline; identical output at shipped defaults
• Cover all gpuVectorE/vignetteVector lanes in the packing test

Review gate

remote review gate selected — GitHub claude-review is the enforced reviewer.

Testing

• Metal golden-image suite compares unchanged (proves zero visual drift at defaults)
• Tuning, entrance suites pass (18 tests)

🤖 Generated with Claude Code

[claude]

Reviewed the diff — one clear issue plus one worth double-checking.

Test file has duplicate assertions (KeyboardStageTuningTests.swift)
The added block re-asserts two lines that already exist in the file:

• XCTAssertEqual(tuning.gpuVectorE.x, tuning.fillLightStrength) appears twice (once as pre-existing context, once newly added).
• XCTAssertEqual(tuning.gpuVectorE.z, tuning.wearStrength) also appears twice (newly added, then again as the original trailing line).

Looks like a copy/paste slip when appending the new .y/.w/vignette assertions — the duplicates don't add coverage and should be dropped, leaving just the new, distinct checks for fillSpecularStrength, highlightJitter, vignetteStrength, and vignetteInnerRadius.

Unbounded wear now feeds unclamped specular gain (KeyboardStage.metal)
Changing saturate(uniforms.tuningE.z) → max(uniforms.tuningE.z, 0.0) intentionally removes the old implicit upper clamp so wearStrength values above 1.0 can actually amplify wear (previously saturate silently capped it at 1.0, which — combined with the old default of 1.5 — meant the multiplier above 1 was a no-op; good catch). However, wear now has no upper bound at all, and it's used unclamped in:

float keySpecularGain = isKey
    ? (1.0 + (materialSeedB - 0.5) * 0.30 * min(uniforms.tuningE.w, 1.0))
        * (1.0 + wear * 0.55)
    : 1.0;

The roughness expression a few lines up is still bounded by the surrounding clamp(..., 0.12, 0.82), but keySpecularGain/specular aren't clamped in the shown diff, so a wearStrength tuned above ~2 could blow out the specular highlight on high-wear keys. If wearStrength is only ever set from a small, developer-controlled range this is fine — but if it's exposed to broader tuning/config, consider clamping wear (e.g. min(wear, someMax)) before it feeds the specular gain.