Initialize Miles LoRA A matrices with kaiming-uniform like the megatron backend

[kevintli]

Summary

The Miles backend never sets args.lora_A_init_method, so Miles' create_multi_lora_instance falls back to getattr(args, "lora_A_init_method", "xavier"). In the bundled Megatron-Bridge MultiLoRA layers, "xavier" means xavier_normal_. Spindle's own megatron LoRA backend (megatron_runtime/lora/model.py) passes lora_A_init_method="kaiming" (kaiming_uniform_(a=sqrt(5)), the PEFT default). Thinking Machines' "LoRA Without Regret" post says Tinker uses that PEFT parametrization: uniform A with scale 1/sqrt(d_in), zero B, alpha 32.

# MilesRuntime._start, after parse_args
args.lora_A_init_method = "kaiming"

Why it matters: B starts at zero, and Adam's per-step update to B is about lr regardless of scale. So the early weight change ΔW = (α/r)·ΔB·A scales with std(A), and xavier gives a higher effective LR than Tinker at the same learning_rate.

Measured std(A) in the exported step-0 adapters for gpt-oss-20b (layer 3 and lm_head). The PEFT reference is 1/sqrt(3·d_in):

target	xavier (before)	kaiming (this PR)	PEFT per-matrix
q/k/v_proj, lm_head (d_in 2880)	0.0262	0.0108	0.0108
o_proj (d_in 4096, TP-sharded)	0.0605	0.0255	0.0090
experts gate_up/down (packed 3D)	0.0044	0.0019	0.0108

This PR fixes the dense q/k/v/lm_head targets exactly. Two gaps remain because the init runs on local, packed tensors, which gives the wrong fan-in:

• Row-parallel o_proj sees the 512-wide TP shard.
• Packed experts [local_experts, rank, in] see rank*in.

Fixing those needs per-expert, global-fan-in init in Megatron-Bridge. That is out of scope here; this PR only makes the Miles path consistent with Spindle's megatron path.

Found while reproducing jasper-lu/sec-search-rl, where Spindle policies collapsed to short trajectories faster than the hosted-Tinker reference.

Link to Devin session: https://modal.devinenterprise.com/sessions/f53cfabb210146de8f0338fe388d7973
Open in Devin Desktop: https://modal.devinenterprise.com/desktop/session/f53cfabb210146de8f0338fe388d7973?variant=devin
Requested by: @kevintli

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access that start with 'Devin'.

• Disable automatic comment, CI, and merge conflict monitoring