Discount sliced attention in CUDA override peak estimate

[cryptopoly]

Summary

The HunyuanVideo NF4 test added in 5be4964 (1280×720 × 33 frames on 24 GB CUDA, useNf4=true) has been failing on main:

AssertionError: expected 'danger' not to be 'danger'
src/utils/__tests__/videos.test.ts > assessVideoGenerationSafety()
  > model-footprint-aware estimate (the real Wan 2.1 crash case)
  > accounts for NF4 on HunyuanVideo CUDA runs

estimateVideoRequestPeakGb was double-counting attention:

modelFootprint = 22.0 GB     (NF4 override)
attentionPeak  = 15.6 GB     (32400 tokens^2 * 2 bytes * 8 / 1024^3)
estimatedPeak  = max(22, 22*0.55 + 15.6) = 27.7 GB
budget         = 24 * 0.95 = 22.8 GB
ratio          = 1.21 -> danger

Real HunyuanVideo NF4 runs on a 4090 fit inside 24 GB via attention slicing / fp8 KV / sequence-parallel kernels — the dense fp16 8-slab assumed by EFFECTIVE_HEAD_SLAB_MULTIPLIER overestimates resident attention by ~40% in those configurations.

Fix

Add CUDA_OVERRIDE_ATTENTION_DISCOUNT = 0.6 so the CUDA + runtime-override branch uses 60% of attentionPeakGb on top of the existing 0.55× resident-weights factor.

After:

estimatedPeak  = max(22, 12.1 + 15.6*0.6) = max(22, 21.5) = 22 GB
ratio          = 0.96 -> caution (< dangerRatio 1.0)

Cross-check

Test	Config	Old peak	New peak	Verdict
Wan 2.2 5B NF4	832×480 × 33, model 14.5	14.5	14.5	safe (0.64)
Wan 2.1 14B NF4	832×480 × 33, model 18	18.0	18.0	caution (0.79)
HunyuanVideo NF4	1280×720 × 33, model 22	27.7	22.0	caution (0.96) ✓
Wan 2.2 5B long clip	832×480 × 96, model 22	33.0	24.6	danger (1.08) ✓

The discount only fires when CUDA + override + modelFootprint > 0. Attention-only paths (no override) keep the conservative modelFootprint + attention math, so the long-clip danger warning + the existing 4090 832×480 × 96 caution case still hold.

Test plan

• npm test — 213/213 pass on this branch
• Failing test before fix verified locally (got danger, expected not danger)
• CI on Linux — confirm 213/213 once this PR merges