refactor(model): one canonical model class, enforced 3-tuple forward …

[mustfaaaa]

…contract

1. Remove the silent LSTM fallback solver_model.py wrapped its model import in try/except and, on ImportError, silently defined an entirely different LSTM-based CalculusSolverModel, so a broken model/transformer.py trained the wrong network with no error. It is now a hard-import re-export; an unimportable model file fails at import time.

2. Consolidate to one tree-based model class model/transformer.py::CalculusSolverModel is canonical: it is what train.py trains, what checkpoints/final/best.pt is a state dict of, and the only class whose forward(src, tgt) matches how beam_search calls it. model/architecture.py::CalculusModel is retired. Beyond the constructor difference (rule_labels= vs num_rules=) its forward() took four positional arguments, so beam_search could only run it via a TypeError-retry shim. Removed its import and load branch from inference/solve.py; tree-keyed checkpoints now load into the canonical class.

3. Document and enforce the 3-tuple return contract Documented at CalculusSolverModel.forward(). Added check_forward_contract(), which train.py now runs before the first training step (~60 ms at the real config). Call-site audit:

• train.py (4 sites), predict.py: already unpacked explicitly.
• inference/beam_search.py: two isinstance(output, tuple) guards replaced with explicit unpacking through _forward(), which raises a clear TypeError on a contract violation.
• inference/solve.py: PklTransformerModel (legacy model/model.pkl loader) returned a single tensor. It now accepts the canonical signature and returns (logits, None, None); it has no rule or verifier head.
• deployment/: uses the ONNX session's named "logits" output, not a torch forward(), so is unaffected.

4. Reconcile beam_search() signature drift No caller passes src_positions / parent_child_pairs; removed from beam_search(). The _call_model helper that caught TypeError and retried with alternative argument shapes is gone -- it masked genuine TypeErrors raised inside forward() as signature mismatches. test_prefix_parity.py's 4-arg-compatibility test is inverted to assert that a non-canonical model now fails loudly.

5. Interface smoke test tests/unit/test_model_interface.py: instantiates the model, runs a dummy forward pass, asserts the 3-tuple and shapes; covers teacher-forced rule ids, the contract checker rejecting single-tensor and mis-shaped outputs, no fallback path in solver_model.py (AST check), the retired module being gone, the legacy pkl model conforming, and beam_search end to end against a real CalculusSolverModel.

Tests: 246 -> 259 passing, 4 xfailed.

Not changed, flagged: model/simple_transformer.py::SimpleCalculusModel (ONNX export) still returns a single tensor. Changing it alters the exported graph's outputs that deployment/onnx_beam_search.py reads by name, so it is left for the ONNX path's owner.

[vercel]

@mustfaaaa is attempting to deploy a commit to the seno-quantum-coder's projects Team on Vercel.

A member of the Team first needs to authorize it.