new architecture

[chaudhryumer]

Architecture pivot: replace tree-encoder model with SimpleCalculusModel

Root cause: tree-encoder + separate RuleHead + StepTracer design plateaued
at ~0.51 Val Loss across 8+ different training configs. Diagnosed two real
bugs (RuleHead pooling only from token 0, rule/decoder circular dependency)
that improved but didn't resolve the plateau. Replaced with a standard
nn.Transformer encoder-decoder, folding rule prediction into the output
sequence as a leading RULE:xxx token instead of a separate classifier head
Also: added integrate operation training data (previously only diff/partial
existed), added RULE:trig_rule/exp_rule/log_rule vocab tokens to stop
Phase 2 trig/exp/log problems being mislabeled under chain_rule, fixed
[BOS] token handling in beam search and output deserialization, fixed
run_pipeline.py reading wrong result dict key.

Result: eval/run_eval.py accuracy improved from 0% to 18.3% overall
(36.2% diff, 43.3% partial). gradient/tangent_line remain unsupported --
documented as separate follow-up work requiring serializer/schema changes."

[vercel]

@chaudhryumer 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.