[CLM 1/2] Checkpoint and typed decision scoring

[xiaoyu-xyz]

CLM is the second model tracked in #9, and it decides differently in a way LAYA does not cover: the engine does not compute embeddings. A frozen Qwen3-8B encoder runs as its own process, and the engine owns everything after it. This is the half that can be checked without a GPU — the checkpoint, the head geometry, and the decision arithmetic.

Core code: 483 lines (config 75, weights 146, scoring 256, lib 6), within the 500-line budget. Tests and the export/oracle scripts are separate.

This is the first of three, stacked:

		core
this PR	checkpoint, geometry, scoring	483
next	the embeddings client	178
then	the request path, clm-run, and the comparison against CLM	392

The checkpoint is converted first

A CLM checkpoint is a torch.save dict, so it is a pickle and no non-Python reader can open it. recipe/clm/native/export_weights.py writes the tensors to safetensors with the head name as a prefix, and keeps cfg, hidden_size, projection_dim and logit_scale in the metadata. The published CLM_v0.1-8B.pt is 75 MB and holds the two heads, not the 8B encoder; the export is 16 tensors and 18.9 M parameters.

The decision

softmax(exp(logit_scale) * cos(state_head(s), action_head(c)) / temperature) over a question's candidates. Both heads are inp → [LayerNorm →] hidden → out with GELU, and both projections are L2-normalised before the dot product.

The three types differ only after the distribution exists, so they share one path: choice is the argmax key, noul is the true entry of a two-candidate distribution, and score is the expected level index. confidence is the top probability minus the mean of the rest, clamped, and 1.0 for a single candidate — the definition src/clm/schema.py uses.

The scale cap

exp(logit_scale) is capped at 100, as heads.py caps it. The published logit_scale is 4.6132, whose exponential is 100.82, so the cap binds — without it every probability is about 0.8 % off, and more where candidates are close.

I had this wrong until a real encoder disagreed with me. The CPU-side oracle could not catch it: head_oracle.py was written from the same reading of the format and made the same mistake, so both sides agreed and the test was green. It took compare_with_reference.py (in the third PR) comparing against CLM's own heads over real Qwen3-8B embeddings.

Checks

With the checkpoint exported:

python recipe/clm/native/export_weights.py CLM_v0.1-8B.pt /tmp/clm-export
python recipe/clm/native/head_oracle.py /tmp/clm-export /tmp/clm-export/head-oracle.json
CLM_EXPORT=/tmp/clm-export cargo test -p omni-clm -- --include-ignored

5 tests: 16 tensor conversion hashes against the export oracle, the inventory against the head configuration, and five decisions against head_oracle.py — an independent NumPy implementation of the same arithmetic — on hash-derived embeddings. Default cargo test needs no checkpoint (2 passed, 3 ignored). fmt and clippy -D warnings clean.