Cut the per-seed WebGL persona to the two renderer buckets that score clean on
FP Pro tampering_ml across seeds (AMD Radeon R9 200 Series and Intel Arc A750),
weighted 70/30, cross-vendor so the fleet isn't one fixed GPU. Every NVIDIA
bucket and the integrated/ancient Intel buckets are penalised, so they're out.
The canvas/WebGL render-image hash turned out to be the dominant tampering_ml
driver, not the attributes, so the render-noise seed (zoom.stealth.fpp.hw_seed)
is now decoupled from the identity seed and drawn from a calibrated clean pool.
Per-seed determinism and per-user diversity are preserved.
Also in this change:
- audio maxChannelCount is stereo-dominant per class (it reflects the output
device, not the GPU; the old tables over-emitted 5.1/7.1 surround)
- route discrete Intel Arc desktop cards to a discrete-GPU class (not integrated)
- condition the whole sampled profile on the exposed GPU class via the sampler's
evidence path, so cores/screen/storage stay coherent with the declared GPU
- apply per-named-font width factors on Windows/macOS so canvas measureText
widths don't collapse to a single value
12/12 seeds clean on tampering_ml (worst 0.29), bot and anti-detect negative,
and the fingerprint stays identical across repeated runs of the same seed.
Covers classify_gpu (decision-table over 28 GPU strings inc. boundary
values for AMD Radeon number ranges), _screen_tier resolution
classification, derive_font_prefs / derive_font_whitelist coherence and
determinism, and the public Forge / sample entry points (locked
identity, key set, type correctness, seed determinism).
One plan deviation: the original plan claimed `AMD FirePro W7100` →
`workstation`, but the workstation regex requires a `Radeon` prefix,
so FirePro alone falls through to the `mid_range` fallback. Test
asserts the actual behaviour.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
feder-cr