feat(gemma4): global decoder layer and its HF golden oracle probes

[FeathBow]

Description

Closes #876

The global (full-attention) decoder layer at head_dim 512, forward only — everything downstream of attention was already kind-agnostic and is now one shared epilogue. The differences from the local kind are the ones the checkpoint and the reference pin:

• K=V is a runtime fork, not shared storage. The checkpoint ships no v_proj on any full_attention layer (the loader enforces it; this layer refuses one that carries it), and the reference forks value off k_proj's raw output BEFORE k_norm and RoPE touch K — after the norms the two tensors differ bitwise.
• Proportional RoPE is table-borne. The inverse frequencies keep the full head_dim as exponent denominator, the un-rotated band is zero-padded, and rotate_half pairs (d, d + 256) across the whole head — so the prep kernel runs at rotary_dim = head_dim and the partiality lives in the tables. Passing the intuitive 128 rotates the wrong pairs entirely; a control below shows the oracle catches exactly that. The rope contract is parsed and validated once on the typed config — each family's declared rope_type is asserted against the implemented algorithm, theta must be positive, and partial_rotary_factor * global_head_dim must land on a positive even width within the head — and the table builder re-checks its inputs at the seam. The global forward range-checks positions on the host before upload, because its batched prep kernel __trap()s on out-of-range positions instead of returning an error (the hd256 prefill wrapper rejects them itself).
• Attention stays unscaled (scaling = 1.0) through the hd512 single-prefill's explicit sm_scale, and the layer is prefill-from-zero only (start_pos > 0 rejected), like the local kind.

The oracle extends to both global boundaries — indices from the layer-map parse, cross-checked against the fixture metadata's own parse — eight comparisons in one checkpoint load. The last global layer's input is the sliding side's output cut: the fixture dedups adjacent cuts by depth, and that adjacency is asserted rather than assumed.

Test Env

Single GPU (sm_89, x86_64), CUDA 12.9, against the pinned 12B checkpoint.

Verification

• Eight comparisons (single/short x sliding/global x first/last boundary) against the declared tolerance of 0.4 absolute + 2% relative:
  • All four one-token comparisons are bitwise exact (max_abs = 0) — including both global layers. With softmax over one key, query, RoPE and scale drop out of the result, so this pins the V/O path — the K=V fork and the weightless V norm included — the norms, the MLP and layer_scalar exactly; the tables and the scale are gated by the nine-token comparisons and the controls below.
  • The four nine-token comparisons show scattered rounding noise only, zero elements over tolerance, token 0 exact throughout; the global layers sit below the local pair (max_abs 0.125 / 0.0625 vs 0.1875 / 0.25).
• Five negative controls, each against the full checkpoint: swapping the reference cut reddens all eight; sm_scale pinned to rsqrt(512) fails exactly the two multi-token global comparisons; the V fork moved after the norms fails both multi-token global comparisons while one-token stays in tolerance (the weightless norm cancels k_norm's magnitude until multi-key softmax amplifies the residue — why the multi-token probe exists); rotary_dim = 128 — the intuitive reading — fails both multi-token global comparisons; rebuilding the tables with rotary_dim as exponent denominator fails the first global comparison.

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update

[FeathBow]

@codex review please

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Already looking forward to the next diff.

Reviewed commit: d033bd1cb9

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".