TriAttention V3 hybrid recipe: two fixes for Qwen3.5 NIAH failure

[CG-8663]

Response to the open question in triattention-v3.md Section 5 — "failure modes to share or a recipe that fixes the hybrid case."

Two Fixes

Fix 1: Scale eviction budget by attention fraction

The root cause: Qwen3.5-27B has 16/64 attention layers. At 90% retention, you are removing 10% of tokens from a model where each attention token does 4x the work. That is equivalent to 40% effective eviction on a full transformer.

Formula:

effective_budget = 1.0 - (1.0 - raw_budget) * attention_fraction

Qwen2.5-7B:  attention_fraction = 32/32 = 1.0 → 90% (unchanged)
Qwen3.5-27B: attention_fraction = 16/64 = 0.25 → 97.5% (evict 2.5%)

Fix 2: Partial RoPE frequency count

Qwen3.5 rotates only 64/256 head dimensions. The scoring loop iterates over head_dim/2 = 128 frequency bins, but 96 of those contribute zero signal (no rotation = no position encoding = no trig score difference). The score averages 32 bins of signal with 96 bins of noise.

Fix: iterate only over n_rot/2 frequencies.

int freq_count = model->hparams.n_rot / 2;  // 32 for Qwen3.5, 64 for standard

Why These Should Work

• Fix 1 explains every observation: PPL fine (Mamba dominates), NIAH fails at mid/end (attention is sparse), start passes (prefix protection saves it)
• Fix 2 is a 4x signal-to-noise improvement on the scoring that chooses which tokens to evict
• Both fixes are derived from the model architecture, not tuning
• Neither changes the scoring formula — V3 trig scoring is correct, it just needs the right inputs

Validation Recipe

Included in the document — step-by-step commands for testing each fix independently and stacked, including TurboQuant+ integration.

TQBridge Integration

With both fixes: TriAttention + TurboQuant on reasoning workloads = ~23x combined compression = 2.2KB per token over the wire. Viable over WiFi for distributed inference.

Full analysis: docs/papers/triattention-hybrid-recipe.md