v18-safe-defaults: flip SUR_LORA_INIT=loftq + DISABLE_AL=1 as defaults

V#7 burned 9 of 9.1 wall-clock hours on AL pre-filter scoring 20K samples
at 1.6s/sample BEFORE training even began, then crashed with PiSSA + 4-bit
quantization incompatibility ('Please initialize PiSSA under fp32/fp16/bf16').

Both failure modes are now avoided WITHOUT requiring the user to remember
to set env vars:

- SUR_LORA_INIT default: pissa_niter_4 → loftq (4-bit-safe)
- DISABLE_AL default: 0 → 1 (skip 9h pre-filter)

Override only if the base loads in fp16/bf16 (e.g. SUR_LORA_INIT=pissa_niter_4
when not running 4-bit), or once a stable adapter exists and we want the AL
teachable-zone filter back on for a refinement run (DISABLE_AL=0).

bin/kaggle-trainer.sh

@@ -1026,7 +1026,11 @@ if USE_LIGER:
 # Cost: 1 fwd pass per scored sample, ~30-60 ms each on T4 7B 4-bit.
 # AL_SAMPLE_CAP=20000 → ~10-20 min budget. Skip with DISABLE_AL=1 or if
 # raw is below the floor (5000 rows — not enough signal to bother).
-DISABLE_AL = os.environ.get("DISABLE_AL", "0") == "1"
+# V18 default flipped to disabled — V#7 spent 9 of 9.1 wall-clock hours
+# scoring 20K samples at 1.6s/sample BEFORE training even started. With
+# Kaggle's 30h/week budget, AL filter is an unsustainable upfront cost.
+# Re-enable explicitly with DISABLE_AL=0 once a stable adapter exists.
+DISABLE_AL = os.environ.get("DISABLE_AL", "1") == "1"
 AL_SAMPLE_CAP = int(os.environ.get("AL_SAMPLE_CAP", "20000"))
 
 if DISABLE_AL or len(raw) < 5000:
@@ -1111,7 +1115,11 @@ lora_kwargs = dict(
 #                     task-aware SVD with quant-awareness in one shot. peft
 #                     ≥0.13 with CordaConfig. Falls back to pissa if missing.
 #   gaussian        — Kaiming default (ablation baseline)
-LORA_INIT = os.environ.get("SUR_LORA_INIT", "pissa_niter_4")
+# V18 default flipped to "loftq" — PiSSA + 4-bit BitsAndBytes crashed V#7 at
+# 9.1h with `Please initialize PiSSA under float32, float16, or bfloat16`.
+# LoftQ is the safe 4-bit-aware path. Override with SUR_LORA_INIT=pissa_niter_4
+# only if the base model is loaded in fp16/bf16 (no 4-bit quant).
+LORA_INIT = os.environ.get("SUR_LORA_INIT", "loftq")
 try:
     from peft import LoraConfig as _Probe
     import inspect