diag(distill): per-kernel JIT logging in forward cache (PMAT-698i)

[noahgift]

Summary

The Phase 3 cuda dispatch defect cascade (PMAT-698e..h) has been chasing missing pre-warm entries ONE KERNEL AT A TIME by inferring kernel names from CUDA error messages. After 5 iterations:

Iter	PR	Kernel discovered	Result
1	#1808 PMAT-698e	(workspace cap, not a kernel)	unblocks load
2	#1809 PMAT-698f	(format compat, not a kernel)	unblocks pipeline
3	#1810 PMAT-698g	silu_backward, softmax_backward, rms_norm_backward	unblocks 3 backwards
4	#1813 PMAT-698h	rms_norm_gamma_reduce	unblocks 1 more
5	❓	Still 4 GH-480 patches post-pre-warm with NO kernel name in log	unknown

This diagnostic PR adds eprintln to ForwardKernelCache::get_or_compile so every forward JIT event prints its kernel key. After one dispatch we get the full list in one pass.

Mirrors the existing [BWD-CACHE] log on the backward path.

Sample expected output (post-merge dispatch)

[CUDA] Pre-warmed 23 forward kernels (JIT compiled before block upload)
...
✓ 24 transformer blocks uploaded to GPU
[FWD-CACHE] Compiling 'batched_rope_neox_forward_896' (ptx_len=4231)
[FWD-CACHE] OK 'batched_rope_neox_forward_896'
[FWD-CACHE] Compiling 'bf16_to_f32_cast' (ptx_len=1108)
...

Then PMAT-698j adds those kernel names to pre_warm_for_model.

Test plan

• cargo check --features cuda clean build
• Live gx10 dispatch logs every forward JIT event (post-merge)
• Each kernel name becomes a one-line addition to pre_warm_for_model

Defect cascade — stage 6 (DIAGNOSTIC)

#	PR	What
1	#1804 PMAT-700-B	Skip PTX GEMM pre-warm (JIT cache pressure)
2	#1808 PMAT-698e	Cap max_seq_len at 2048
3	#1809 PMAT-698f	Accept APR magic in weights loader
4	#1810 PMAT-698g	Pre-warm non-LoRA backward kernels
5	#1813 PMAT-698h	Pre-warm rms_norm_gamma_reduce
6	THIS PMAT-698i	Diagnostic: log every forward JIT
7	(next) PMAT-698j	Consolidated pre-warm sweep from diagnostic output

🤖 Generated with Claude Code