feat(tune): surface-B GDN LoRA weight gradients + train_grad_full fields

[ohdearquant]

What

Surface-B: LoRA weight gradients through GDN (gated delta-net) layers, plus the train_grad_full field fixes and a forward-parity test.

Why

Extends the micro-LoRA backward pass to compute weight gradients for GDN-layer LoRA adapters (previously only attention/FFN surfaces had them). The Option-gated None path is byte-identical to before — when GDN LoRA grads are not requested, behavior is unchanged.

Files

• crates/inference/src/attention/gdn_backward.rs (+978/-29, cfg-gated under train-backward)
• crates/tune/src/bin/train_grad_full.rs (+592/-132, includes the E0063 GDN-fields fix)
• crates/tune/src/bin/train_grad_layer23.rs
• crates/inference/src/backward/tape.rs, crates/inference/src/backward/ops.rs
• crates/inference/examples/diff_gdn_layer.rs (required-features = train-backward,f16)
• crates/inference/tests/lora_forward_parity_test.rs (+917, new; compiles under default features via internal cfg guards, CI-safe)

Verification

cargo build --release -p lattice-tune --bin train_grad_full --features train-backward clean. The train_grad_full E0063 (missing GDN gradient fields) was the second of the two regressions that motivated the PR-G gate. Built green in the integrated-tree gate.

⚠️ Gradcheck PENDING — do not merge yet

The numerical gradient-check for the GDN weight-grad surface has not been run (tracked: khive task fe42740c). The forward path and compile are verified; the backward correctness is unverified. This PR is up for review structure and to keep the code on a branch, but the gradcheck must pass before merge. Marking ready-for-review on the non-gradient parts; holding the GDN-grad merge on that run.

Bench

Backward/training code is train-backward-gated, off the default decode path. make bench-compare's comparator errored assembling the delta (known two-worktree fragility; base benches ran clean) — bench-neutral by construction.

Series

Part of the PR #193 engine-slice (finest split). All engine code lands on main; the macOS app surfaces a subset (Models + Chat) for v0.0.1.

[github-actions]

E2E Parity Report

PASS: all 3 prompts match within first 3 tokens

Prompt	Agreement	First Diff	HF tok/s	Lattice tok/s	Verdict
The capital of France is	3/15	pos 3	0.4	2.0	PASS
In the year 2024, artificial intelligence	10/15	pos 9	0.4	1.9	PASS
`def fibonacci(n):

if n <= 1:
    return n
return` | 15/15 | none | 0.3 | 1.7 | PASS |

The capital of France is

• HF: Paris.
  The capital of France is Paris.
  The capital of France
• Lattice: Paris.
  A: Yes, the capital of France is Paris.

In the year 2024, artificial intelligence

• HF: (AI) has become a significant part of the global economy. It is
• Lattice: (AI) has become a significant part of our daily lives. From personal

def fibonacci(n): if n <= 1: return n return

• HF: fibonacci(n-1) + fibonacci(n-2)

print(fib

• Lattice: fibonacci(n-1) + fibonacci(n-2)

print(fib