fix(gen-seat): AEON W4A4 was the defect — purged; mixed FP8-attn build is primary gen

Root cause of the multi-day degeneration hunt, operator-confirmed: the AEON
NVFP4 W4A4 quant (sakamakismile/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-NVFP4,
full W4A4 incl. attention) went degenerate ~15-20% of generations in real
multi-turn use and forced regenerates. MTP, prefix-caching, and the gateway
all merely AMPLIFIED it, which is why MTP-off, APC-off, and the vLLM #51113
fix each 'helped' a synthetic probe without fixing it -- three plausible
false root-causes, each passing one clean run then failing in real use.

The fix was the WEIGHTS: the in-house JonathanColetti/Heretic mixed
NVFP4+FP8 build (qwen38-27b-uncensored-nvfp4-mixed, FP8 attention not W4A4,
same base, same MTP) is coherent through long multi-turn with MTP ON. W4A4
*attention* was the defect; FP8 attention is not.

This commit:
- GEN_MODEL -> the mixed FP8-attn build (primary gen until DavidAU 3.8 lands)
- GEN_IMAGE pinned to vllm/vllm-openai:nightly-311b3513... (v0.27.2rc1.dev150,
  carries #51113; pinned by sha so it does not drift on the next pull)
- AEON weights PURGED from /tank (no-good), safety-checked not-in-use first
- playbook 3.8: the stochastic-W4A4-degeneration lesson + isolate-weights-early
  + do-not-declare-a-fix-from-one-probe (it validated three non-fixes)

AEON is re-pullable from HF if ever needed, but the operator ruled it no-good.

docs/pfi/model-quantization-playbook.md

@@ -131,6 +131,41 @@ There is also a **modelopt-format-specific** version of this: vLLM 0.24 does not
 modelopt `exclude_modules` to the spec-decode *draft* model, which no checkpoint config can fix
 (needs a `sitecustomize` runtime patch). Using compressed-tensors avoids it entirely — §3.4.
 
+### 3.8 ⭐⭐ A bad W4A4 quant degenerates STOCHASTICALLY — and masquerades as everything else
+
+The single most expensive misdiagnosis this project has had. A **defective NVFP4
+W4A4 quant** (`sakamakismile/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-NVFP4`, full
+W4A4 incl. attention) went degenerate **~15-20% of generations** in real
+multi-turn use — collapse, cross-turn content bleed, early-ending — forcing a
+regenerate. It was **purged 2026-08-17** as no-good.
+
+The reason it cost days: **every other layer AMPLIFIED it, so each looked like the
+cause in turn.** MTP made it worse (so MTP-off "fixed" a synthetic probe); prefix
+caching made it worse (so APC-off "fixed" another probe); the vLLM #51113 mamba
+fix made it "significantly better." Each partial-improvement produced a plausible
+false root-cause that passed a clean run and then failed in real use. The actual
+cause was the **weights**: swapping to the FP8-attention mixed build
+(`qwen38-27b-uncensored-nvfp4-mixed`, same base model, same MTP, same nightly, MTP
+ON) is **coherent**. W4A4 *attention* was the defect; FP8 attention is not.
+
+**Lessons, dearly bought:**
+1. **Stochastic degeneration (15-20%) is nearly invisible to a small synthetic
+   probe.** A single 7-turn run passes ~4 times in 5. n=1 "clean" proves nothing;
+   this class of bug needs many runs or the operator's real high-volume use. Do
+   NOT declare a fix from one passing probe — it validated three non-fixes here.
+2. **Isolate the WEIGHTS early.** When a quant degenerates, swap to a *different
+   quant of the same base* before chasing serving flags (MTP, prefix-caching,
+   engine version). Had the AEON→mixed weight swap come first, the serving-flag
+   rabbit holes (MTP, APC, SGLang, nightly) would have been unnecessary.
+3. **W4A4 on attention is high-risk for coherence, independent of the acceptance
+   gate.** AEON passed every static gate — abliteration 4/4, surface 6/6, a 36k
+   needle, 52% MTP acceptance — and was still stochastically broken in
+   conversation. Prefer FP8 attention (the mixed recipe, §2) over full W4A4; if
+   W4A4 attention is used, gate on *many* real multi-turn generations.
+
+Current primary gen: the mixed FP8-attention build on vLLM nightly with MTP,
+until the DavidAU Qwen3.8 lands.
+
 ### 3.7 ⭐ A LOADED MTP head can still corrupt output — Qwen3.8 multi-turn
 
 §3.3 is about *losing* the head (0% acceptance, silent). This is the opposite and

stacks/gen-seat/.env.example

@@ -1,5 +1,5 @@
 # gen-seat tunables — fleet `gen` seat (ana-ml2 GPU 0, :8015). Edit here, never commit.
-GEN_IMAGE=vllm/vllm-openai:latest
+GEN_IMAGE=vllm/vllm-openai:nightly-311b3513af33bc29b4acb2fde2e9313e5e9966a0
 API_KEY=
 GEN_GPU_ID=0
 

stacks/gen-seat/compose.yaml

@@ -126,10 +126,18 @@ services:
       # in a stable release — so we do not jump the fleet gateway to an RC.
       # Cost of MTP-off: ~half decode tok/s. Accepted as the known-good.
       #
-      # TO RE-ENABLE (only with real-use verification): uncomment the two lines
-      # below. Do NOT trust a synthetic probe alone — it passed and was wrong.
-      # - --speculative-config
-      # - '{"method": "${GEN_SPEC_METHOD:-qwen3_5_mtp}", "num_speculative_tokens": ${GEN_SPEC_TOKENS:-3}}'
+      # MTP ON — PRIMARY GEN CONFIG (resolved 2026-08-17, operator-confirmed
+      # coherent). The whole multi-day degeneration hunt root-caused to the AEON
+      # W4A4 quant being DEFECTIVE: it went degenerate ~15-20% of generations and
+      # forced a regenerate. MTP / prefix-caching / gateway all merely AMPLIFIED
+      # it, which is why MTP-off and the vLLM #51113 fix each "helped" without
+      # fixing it. This build — the in-house JonathanColetti/Heretic mixed
+      # NVFP4+FP8 weights (FP8 attention, NOT W4A4) on vLLM nightly (#51113 fix)
+      # with MTP ON — is coherent through long multi-turn in real use. AEON
+      # (qwen38-27b-aeon-ultimate-nvfp4) was PURGED as no-good. Primary gen until
+      # the DavidAU Qwen3.8 lands.
+      - --speculative-config
+      - '{"method": "${GEN_SPEC_METHOD:-qwen3_5_mtp}", "num_speculative_tokens": ${GEN_SPEC_TOKENS:-3}}'
     deploy:
       resources:
         reservations: