v18: add Qwen3.5/3.6 aliases (2026-04 releases) — Qwen3.5-9B = sweet spot for T4x2

HF API audit 2026-04-30 confirmed Qwen3.5 (Apr-23) and Qwen3.6 (Apr-24)
families are live. User flagged that earlier base-swap recommendations
defaulted on Qwen2.5 without surfacing the Apr-2026 options.

Reality on T4x2 16GB/card (V5 OOM trace shows 14B already at the edge):
- Qwen3.6-35B-A3B / Qwen3.6-27B → tight to OOM, MoE/dense both ≥13.5GB
- Qwen3.5-35B-A3B / Qwen3.5-27B → same range
- Qwen3.5-9B (dense) → ~4.5GB at 4-bit, comfortable
- Qwen3.5-4B (dense) → trivial

New aliases (set via BASE_MODEL env):
- qwen3.5-4b / qwen3.5-9b / qwen3.5-27b / qwen3.5-27b-int4 / qwen3.5-35b-a3b
- qwen3.5-122b-a10b / qwen3.5-397b-a17b
- qwen3.6-27b / qwen3.6-27b-fp8 / qwen3.6-35b-a3b / qwen3.6-35b-fp8
- qwen3-coder-next / qwen3-coder-480b

Auto-pick stays on Qwen2.5-Coder-7B-Instruct because the existing
axentx/surrogate-1-* HF datasets were distilled with that tokenizer;
flipping auto-default would silently re-tokenize and invalidate the v1
baseline comparison. Override is a one-line BASE_MODEL=qwen3.5-9b.

bin/kaggle-trainer.sh

@@ -416,6 +416,13 @@ def pick_base_for_hardware():
     #   32B 4-bit (16GB) — needs per-GPU ≥30GB (forward+backward eats 2× weights)
     #   14B 4-bit  (7GB) — needs per-GPU ≥22GB safely (V5 OOM proof at 16GB)
     #    7B 4-bit (3.5GB) — fits T4 16GB with margin (~12GB peak)
+    # V18 update (2026-04-30): Qwen3.5-9B (released 2026-04-23) is the
+    # newest dense Qwen that fits T4x2 16GB/card at 4-bit (~4.5GB weights).
+    # Qwen3.6 family is 27B+ — too big for T4x2. Auto-pick stays on
+    # Qwen2.5-Coder-7B-Instruct because the existing axentx/surrogate-1-*
+    # corpus was distilled with that tokenizer; switching auto-default
+    # would silently re-tokenize all data and invalidate v1 baseline.
+    # Set BASE_MODEL=qwen3.5-9b explicitly for the new-model test path.
     if per_gpu_gb >= 30:
         return "Qwen/Qwen2.5-Coder-32B-Instruct", 32.0
     if per_gpu_gb >= 22:
@@ -425,24 +432,45 @@ def pick_base_for_hardware():
 
 _auto_base, _auto_size = pick_base_for_hardware()
 _BASE_ALIASES = {
-    # short-name → real HF path (matches BASE_SWAP_CANDIDATES dict below;
-    # resolved here so BASE_MODEL=qwen-coder-7b just works without users
-    # having to type the full HF repo path).
-    "qwen-coder-7b":      "Qwen/Qwen2.5-Coder-7B-Instruct",
+    # short-name → real HF path. BASE_MODEL accepts either the alias OR the
+    # full HF repo path; aliases are resolved here, anything else passes
+    # through unchanged. Audit dates: 2026-04-30 HF API sweep.
+    #
+    # Qwen2.5-Coder family (current corpus tokenizer match)
+    "qwen-coder-7b":      "Qwen/Qwen2.5-Coder-7B-Instruct",   # Kaggle default
     "qwen-coder-14b":     "Qwen/Qwen2.5-Coder-14B-Instruct",
     "qwen-coder-32b":     "Qwen/Qwen2.5-Coder-32B-Instruct",
-    "granite-4.1-8b":     "ibm-granite/granite-4.1-8B-base",
-    "olmoe-1b-7b":        "allenai/OLMoE-1B-7B-0924-Instruct",
-    "qwen3-coder-7b":     "Qwen/Qwen3-Coder-7B-Instruct",
-    "qwen3-coder-30b":    "Qwen/Qwen3-Coder-30B-A3B-Instruct",
+    # Qwen3 dense Instruct (250k vocab, native Thai BPE)
     "qwen3-7b-instruct":  "Qwen/Qwen3-7B-Instruct",
     "qwen3-8b-instruct":  "Qwen/Qwen3-8B-Instruct",
+    # Qwen3-Coder family
+    "qwen3-coder-7b":     "Qwen/Qwen3-Coder-7B-Instruct",
+    "qwen3-coder-30b":    "Qwen/Qwen3-Coder-30B-A3B-Instruct",  # MoE 3B-active
+    "qwen3-coder-next":   "Qwen/Qwen3-Coder-Next",               # 80B-A3B Next
+    "qwen3-coder-480b":   "Qwen/Qwen3-Coder-480B-A35B-Instruct", # frontier MoE
+    # Qwen3.5 series (2026-04-23 release, "newest that fits T4x2")
+    "qwen3.5-4b":         "Qwen/Qwen3.5-4B",                    # ✅ T4x2 trivial
+    "qwen3.5-9b":         "Qwen/Qwen3.5-9B",                    # ✅ T4x2 sweet spot
+    "qwen3.5-27b":        "Qwen/Qwen3.5-27B",                   # ⚠️ T4x2 risky
+    "qwen3.5-27b-int4":   "Qwen/Qwen3.5-27B-GPTQ-Int4",         # pre-quant T4x2 OK
+    "qwen3.5-35b-a3b":    "Qwen/Qwen3.5-35B-A3B",               # MoE → L40S+
+    "qwen3.5-122b-a10b":  "Qwen/Qwen3.5-122B-A10B",             # MoE → H100
+    "qwen3.5-397b-a17b":  "Qwen/Qwen3.5-397B-A17B",             # frontier MoE
+    # Qwen3.6 series (2026-04-24 release, NEWEST — only 27B+ available)
+    "qwen3.6-27b":        "Qwen/Qwen3.6-27B",                   # ⚠️ T4x2 risky
+    "qwen3.6-27b-fp8":    "Qwen/Qwen3.6-27B-FP8",               # FP8 inference
+    "qwen3.6-35b-a3b":    "Qwen/Qwen3.6-35B-A3B",               # MoE → L40S+
+    "qwen3.6-35b-fp8":    "Qwen/Qwen3.6-35B-A3B-FP8",
+    # Other 7-9B class
+    "granite-4.1-8b":     "ibm-granite/granite-4.1-8B-base",
+    "olmoe-1b-7b":        "allenai/OLMoE-1B-7B-0924-Instruct",
+    # GLM family (template/tokenizer validation toward V19)
     "glm-4-9b-chat":      "zai-org/glm-4-9b-chat",
     "glm-4-9b-chat-1m":   "zai-org/glm-4-9b-chat-1m",
     "glm-4.1v-9b-think":  "zai-org/GLM-4.1V-9B-Thinking",
     "glm-4.7-flash":      "zai-org/GLM-4.7-Flash",
-    "glm-4.5-air":        "zai-org/GLM-4.5-Air-Base",
-    "glm-5":              "zai-org/GLM-5",
+    "glm-4.5-air":        "zai-org/GLM-4.5-Air-Base",          # ❌ T4x2
+    "glm-5":              "zai-org/GLM-5",                     # ❌ V19 only
     "glm-5.1":            "zai-org/GLM-5.1",
     "glm-5.1-fp8":        "zai-org/GLM-5.1-FP8",
 }