Gate the fitted layout score on measured execution classes (calibration expansion, stage 1)

[bradhilton]

Stage 1 of the expanded cost-model calibration campaign (other hidden sizes, MoE), per the plan agreed with Codex: no new GPU evidence yet, only the machinery so that new execution classes can be measured, fitted and certified without loosening the current gate.

What changes

Admission by execution class, not by hidden size. A calibrated coefficient table applies only to the execution classes it was measured on:

• ModelGeometry, read from the Megatron config (never a model name): hidden and FFN widths, attention head geometry, GDN state shape, expert geometry. Layer counts remain scoring facts, not identity.
• DeviceClass: compute capability plus a memory-system bucket (the 80 GB H100 shares capability 9.0 with the H200 but is not admitted).
• ParallelShape: TP × CP × EP × ETP.
• CalibratedTable + select_scoring replace CalibrationProfile / coefficient_version_for. The shipped dense hidden-2,560 table admits exactly the Qwen3.5-4B and Qwen3-4B geometries at the four measured shapes (TP1 × CP1/2/4, TP2 × CP1). Consequently TP2 × CP2, CP8, TP4 and any expert parallelism now fall back to the version-1 score until measured (previously TP2 × CP2 would have used the fitted table unmeasured). CPU-only planning keeps the default table.
• The table identity joins the coefficient version in the planner facts, the layout cache key and the memory-profile signature; the selector threads the table through the scorer.

Runtime facts. TrainerRank derives the geometry, EP/ETP and the MoE layer count from the runtime. Both admitted geometries were cross-checked against providers built on CPU for the two models (identical).

Harness. Cell rows record the geometry, device memory class, EP/ETP, MoE facts and the score the runtime actually used (the earlier rows recorded the module constant, which was wrong for fallback runtimes). New cell family cal-ellavox-qwen3 uses the Qwen3-tokenized Ellavox corpus Codex pointed at (git-ignored, SHA-pinned like the Qwen3.5 corpus; 44 groups, 261 histories, max id 151,668), with an explicit tokenizer-agreement check (vocabulary size and a decoded sample) that fails a cell loudly on mismatch.

Fitter. Cell keys carry |ep{n}/|etp{n} only when above one (existing keys unchanged); certificates name their table (--table-id) and record the admitted device classes, dtypes, geometries and shapes; per-model, per-shape and per-family gate reports, judged per group and never pooled, as Codex asked.

Certificate. Regenerated from the same 58-cell evidence with the class facts backfilled: identical table hash. The certificate test now asserts the admitted sets equal the production table's, that every geometry was measured at every admitted shape, and one geometry per model.

Checks

uv run prek run --all-files green; tests/acceptance/trainer_rank_planner + tests/unit/test_trainer_rank*.py + tests/unit/test_planner_cost*.py: 588 passed, 17 skipped (including the opt-in full refit). No GPU runs in this PR.

Next stages

2. Qwen3.5-35B-A3B smoke, then the deconfounded TP1 lattice (CP1/EP1, CP2/EP1, CP2/EP2, CP4/EP1, CP4/EP2, CP4/EP4), the TP2/ETP1 pair and Ellavox groups.
3. Dense controls (Qwen3-1.7B, Qwen3-8B, Qwen3.5-27B, Qwen3-14B) at CP1/CP2/CP4/TP2, plus TP2 × CP2 for the two 4B models.
4. Per-class certificates.
5. Attention-MoE (Qwen3-30B-A3B) and the normalized cross-model scorer as an offline leave-one-model-out study.

🤖 Generated with Claude Code