…LAS) — revert Bug B's slow path (PMAT-704)

Post-mortem of the PMAT-701 cascade revealed PR #1869 (Bug B) routed Q4K
teachers to `RealizarQ4KTeacher`, a CPU-heavy forward path (layer-norm +
attention + softmax all on CPU; only Q4K matmuls dispatch to GPU). The
7B vocab-aligned 500-step validation hung at step 0 for 1.5 h with GPU
at 0% utilization — empirical proof the path is unusable as a real
distillation teacher.

Five-whys (evidence/distill-7b-cublas-cudatrainer/findings.json):

* Why 1: 7B 500-step validation hung at step 0 → RealizarQ4KTeacher's
  forward runs most ops on CPU; GPU stayed at 0%.
* Why 2: realizar's `OwnedQuantizedModelCuda::forward_cuda` is mostly
  CPU SIMD; only individual Q4K matmuls dispatch to GPU.
* Why 3: PR #1869 picked that path to avoid an F32 dequant at upload
  (claimed "28 GB inflation + student → Linux OOM-kill").
* Why 4: That claim was based on a single SIGKILL observation with
  MANAGED_MEMORY=1 explicit, never verified via dmesg as actual OOM-killer,
  never re-tested under PMAT-701 Bug A's autodetect default.
* Why 5: Cascade momentum + incomplete root-cause discipline. The cheap
  experiment (one-line dispatch flip) would have rejected the hypothesis
  in 5 minutes; a multi-PR architectural detour shipped instead.

Root cause: conflated the cuMemAlloc 30 GB ceiling (real, fixed by Bug A)
with a step-0 SIGKILL on the explicit-managed path (phantom, never
verified). The F32 dequant fits in 128 GB unified memory; cuBLAS is the
fast path and the right default.

## Fix

`crates/apr-cli/src/commands/distill.rs::run_cuda_backend`:

* New env var `APR_DISTILL_TEACHER_BACKEND` with values:
  - `auto` (default) → CudaTrainerTeacher (cuBLAS, F32 dequant)
  - `cudatrainer` → CudaTrainerTeacher (explicit)
  - `realizar-q4k` → RealizarQ4KTeacher (memory-constrained-device fallback)
* Q4K detection still happens, but only controls the fallback path's
  availability — the DEFAULT dispatch is cuBLAS for all teacher types.
* Generic `TruncatingTeacher` wrapper applies PMAT-703 vocab alignment
  uniformly to both backends. Replaces the per-backend truncation that
  PR #1877 was planning to add inside `RealizarQ4KTeacher::from_apr_path_with_target_vocab`.

`contracts/apr-distill-teacher-backend-selection-v1.yaml`:
* 3 equations (backend_dispatch, forward_latency_invariant, bug_b_demotion)
* 4 falsifiers: default routes to CudaTrainer; env override reaches Realizar;
  7B 500-step completes < 30 min; forward parity within Q4K noise floor
* 2 Kani harnesses; qa_gate F-BACKEND-SELECT-001
* Validates clean: `pv validate` reports 0 errors, 0 warnings.

## Verification on gx10

Dispatch log shows the new path firing:

  [PMAT-704] backend=auto → CudaTrainerTeacher (cuBLAS) [override with APR_DISTILL_TEACHER_BACKEND=realizar-q4k for memory-constrained dGPU]
  [PMAT-703] vocab alignment: teacher native=152064, student=151936 → truncating teacher logits to student vocab
  [PMAT-704] teacher backend = CudaTrainerTeacher [Q4K/Q6K (dequant to F32 at GPU upload; cuBLAS GEMM)]
  [CUDA] cuBLAS initialized — forward TF32 tensor cores (41x vs SIMD)
  ✓ Loaded pre-trained weights successfully (APR)
  ✓ 28 transformer blocks uploaded to GPU

GPU utilization observed at 96% during training (was 0% on the
RealizarQ4KTeacher path). cuBLAS is dispatching correctly; the cascade
is no longer hanging at the first step.

## Cascade context

This is the fifth fix in the PMAT-701 family:
  - #1863 Bug A: allocator autodetect Grace Blackwell
  - #1869 Bug B: RealizarQ4KTeacher (now demoted to opt-in fallback)
  - #1874 Defect 3 / PMAT-702: apr eval no-fake-pass on broken models
  - #1877 Bug B's vocab alignment (superseded by TruncatingTeacher in this PR)
  - This PR: cuBLAS default + opt-in Realizar fallback (PMAT-704)

The original Bug B contract (cuda-q4k-frozen-teacher-v1.yaml) is
**demoted, not retracted**: its math is correct as a memory-constrained
fallback path; its DEFAULT-PATH claim was wrong on unified-memory
devices.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>