[MatthewHines]

Summary

Three kernels still emit the (d, sum) pair of a q8_1 activation block with a raw make_half2(d, sum), bypassing the finite helper the rest of the tree has used since 19de7f4 / f6601bf (include/strata/kernels/q8_1_finite.hpp):

site	path
fused_gr.cu — gr_q8_tail	fused gate-reduce tail (QFUSE)
verify_kernels.cu — gdn_q8_1_store	batched GDN verify store (QFUSE)
iq_kernels.cu — s26_swiglu_q8_1_kernel	S26 swiglu quantize

A block whose amax overflows (or that reads a non-finite input) emits d = inf/NaN; fp16 turns the pair into NaN (or the pathological (0, 0)). Nothing downstream tests for finiteness — the dot kernels assume the producer upheld the invariant, because every other producer shares q8_1_ds(). A NaN scale poisons the fp16 dot of the quantized activation and every sum downstream of it through the layer stack; a fully non-finite logit row ends at the sampler's all-non-finite fallback — vocabulary id 0, the ! face of #606.

This is the same defect class already fixed at the other emit sites; these three were missed. The fix is the same one-liner pattern: route the pair through q8_1_ds(), which clamps both halves into the fp16-finite range (finite d, non-negative finite sum — the sum is a mean of |x|) and produces for the GPU paths the same bits the CPU pool and the already-clamped GPU sites produce for the same block. Content-blind producer-side clamp; nothing model-text-aware anywhere in it.

Verification

• nvcc -std=c++20 -arch=sm_120 -c clean on all three changed translation units (CUDA 13.4.2).
• Field evidence from a Qwen3-Next-class hybrid deployment (GDN/PLE + sparse experts, conversation cache + speculative decoding): the id-0 (!) degeneration face of After a 36,689-token request at 155k, every later request answers one repeated token until the engine reloads #606 appears on builds where the fused/S26 paths are enabled and is absent on boots that use the clamped path exclusively — consistent with the site inventory. The wider carried-state failure (restart-cures-it, spans sessions) is analyzed in docs/DEGENERATION_RCA.md; it involves two further purity violations of the state cache that this PR deliberately does not change code for.

Scope

3 clamp lines + 2 includes + 2 docs. No behavior, tuning, or policy changes. Deployment-side mitigations we use in the interim (freezing the expert migrator with the existing flag, disabling the suffix-draft source, serve-layer repetition detectors) are listed as prose in the RCA doc so maintainers can route them properly — no code for them here.

Docs added:

• docs/Q8_1_FINITE_DS.md — the invariant, the missed sites, diagnostic criteria for this defect class.
• docs/DEGENERATION_RCA.md — the full root-cause analysis: the restart-cure and cross-session-persistence localization levers, the three purity violations of state = f(prefix), the falsification tests that pinned each one, a field diagnostic protocol, and recommended engine-side remedies (tier-aware checkpoint identity for the residency split; reproducible policy inputs for the drafting side).

[Niko1221]

Thank you, taken. The three emit sites now go through q8_1_ds(). On top of your change the three sites also use q8_1_finite(d) and q8_1_quant() for the scale and the int8 values, as native_quantize_q8_1_kernel, the swiglu quantizer and the CPU pool do, so a block that overflows stores values that match its stored scale. A block whose amax is finite and below 8.3M is stored bit for bit as before. Identity gates (Q2_0 on an RTX 5070, IQ3_XXS on an RTX 3060, also with STRATA_QFUSE=1 STRATA_LFUSE=1) are identical. The two docs files are not part of this port; the RCA doc makes engine claims we have not checked.

Closing as ported.