Phase 6.4.25b: gate 1 on v2b-bf16 -- full PASS on all buckets

Global weighted loss 0.2725 (vs baseline 0.2750, previous v2-bf16 was
0.2734). All 6/6 non-penpot buckets pass the 0.10 regression threshold,
including delegacion_subagentes which had narrowly failed on v2-bf16
(+0.1019 -> now +0.0882). The 13 corrective seeds didn't hurt forgetting.

Also points vllm-eval at the new v2b-bf16 checkpoint for the gate 2/5
re-measurement.

data/gate1_v2b-bf16_full.log

@@ -0,0 +1,47 @@
+/usr/local/lib/python3.12/dist-packages/torch/library.py:356: UserWarning: Warning only once for all operators,  other operators may also be overridden.
+  Overriding a previously registered kernel for the same operator and the same dispatch key
+  operator: flash_attn::_flash_attn_backward(Tensor dout, Tensor q, Tensor k, Tensor v, Tensor out, Tensor softmax_lse, Tensor(a6!)? dq, Tensor(a7!)? dk, Tensor(a8!)? dv, float dropout_p, float softmax_scale, bool causal, SymInt window_size_left, SymInt window_size_right, float softcap, Tensor? alibi_slopes, bool deterministic, Tensor? rng_state=None) -> Tensor
+    registered at /usr/local/lib/python3.12/dist-packages/torch/_library/custom_ops.py:922
+  dispatch key: ADInplaceOrView
+  previous kernel: no debug info
+       new kernel: registered at /usr/local/lib/python3.12/dist-packages/torch/_library/custom_ops.py:922 (Triggered internally at /opt/pytorch/pytorch/aten/src/ATen/core/dispatch/OperatorEntry.cpp:208.)
+  self.m.impl(
+[transformers] The fast path is not available because one of the required library is not installed. Falling back to torch implementation. To install follow https://github.com/fla-org/flash-linear-attention#installation and https://github.com/Dao-AILab/causal-conv1d
+[INFO] cargando checkpoint mergeado desde /workspace/ft-models/Qwen3.6-35B-A3B-mcp-v2b-bf16
+␍Loading weights:   0%|          | 0/693 [00:00<?, ?it/s]␍Loading weights:  66%|██████▌   | 457/693 [00:00<00:00, 4566.97it/s]␍Loading weights: 100%|██████████| 693/693 [00:00<00:00, 6103.32it/s]
+[INFO] modelo cargado en 298.1s
+[INFO] 147 ejemplos en data/eval.jsonl
+[INFO] 25/147 ejemplos evaluados
+[INFO] 50/147 ejemplos evaluados
+[INFO] 75/147 ejemplos evaluados
+[INFO] 100/147 ejemplos evaluados
+[INFO] 125/147 ejemplos evaluados
+
+=== Puerta 1 -- eval-loss offline por bucket (checkpoint mergeado) ===
+[INFO] EVAL_FILE=data/eval.jsonl
+[INFO] BASELINE_EVAL_LOSS=0.2750
+[INFO] tiempo de eval: 138.7s, memoria pico: 69.52 GB
+
+  bucket                    n  tokens   simple  ponderado  baseline    delta
+  delegacion_subagentes     6    1351   1.3580     1.1657    1.0775  +0.0882
+  manejo_errores            6    1691   0.6403     0.6350    0.5579  +0.0771
+  negativos                 7     920   1.7187     1.7455    1.7138  +0.0317
+  otros_mcps               29    5614   0.3849     0.4088    0.3973  +0.0115
+  penpot                   11    2075   0.8814     0.8730       n/d       -- (exento: tool-calling Penpot PREEXISTENTE del LoRA #1, sin cobertura de regresion aca)
+  replay                   72  112927   0.1807     0.1833    0.1733  +0.0100
+  skills_adherencia        16    4271   1.4482     1.4177    1.3304  +0.0873
+
+  bucket=replay (aislado)  n=  72 loss_avg_simple=0.1807 loss_avg_weighted=0.1833
+
+  loss_avg GLOBAL simple (por ejemplo)     = 0.5514
+  loss_avg GLOBAL ponderado (por token)     = 0.2725
+  baseline (ponderado por token)           = 0.2750
+  diferencia absoluta (ponderado vs baseline) = 0.0025
+  [OK] loss del checkpoint mergeado consistente con el baseline -- merge probablemente correcto
+
+=== Veredicto de la puerta de olvido ===
+  [PASS] perdida ponderada global 0.2725 <= 0.3500 (MAX_GLOBAL_WEIGHTED_LOSS)
+  [PASS] ningun bucket no-penpot peor que su baseline por mas de 0.1000 (MAX_BUCKET_REGRESSION) -- umbral evaluado sobre 6/7 buckets
+
+  VEREDICTO PUERTA 1: PASS
+EXIT_CODE:0

docker-compose.eval.yml

@@ -7,13 +7,15 @@ services:
     ports:
       - "8001:8000"
     volumes:
-      # Fase 6.4.25: apunta al candidato v2-bf16 (base + LoRA #1 + LoRA #2), no al
+      # Fase 6.4.25b: apunta al candidato v2b-bf16 (base + LoRA #1 + LoRA #2
-      # bf16 de Fase 4 (que solo tiene el LoRA #1). Es el checkpoint contra el que
+      # corregido con los 13 seeds del diagnostico de causa raiz), no al v2-bf16
-      # corren las puertas 5 y 2 del disparador de fallback antes de cuantizar.
+      # original (que fallo la puerta 5 modo full por el bug de formato de
-      - /home/aleleba/ft-models/Qwen3.6-35B-A3B-mcp-v2-bf16:/model:ro
+      # tool-call). Es el checkpoint contra el que se remide la puerta 5 tras la
+      # iteracion correctiva, antes de decidir NVFP4 vs fallback.
+      - /home/aleleba/ft-models/Qwen3.6-35B-A3B-mcp-v2b-bf16:/model:ro
     command:
       - "--model=/model"
-      - "--served-model-name=qwen3.6-35b-a3b-mcp-v2-bf16"
+      - "--served-model-name=qwen3.6-35b-a3b-mcp-v2b-bf16"
       - "--tensor-parallel-size=1"
       - "--max-model-len=32768"
       - "--enable-auto-tool-choice"