judge: callable relation models and configurable training pipeline

[karthikreddy-02]

Equation-pair scoring and judge training need reusable Python interfaces for notebooks and representation experiments. This PR adds the trained-judge library, CLI, training scripts, and a configuration-driven pipeline.

Load a Judge once, then call judge(a, b) or compare_many(pairs). Verdicts contain the relation, normalized label scores, and original equations. The loaded PEFT model remains accessible for activation hooks. Labels describe B relative to A.

The pipeline exposes run(config, stage=..., dry_run=...) and an equivalent CLI. Configuration controls rendering, prompts, LoRA settings, checkpoint retention, and evaluation metrics. It supports infix, Lean, LaTeX, Python lambda, and function-call equation notation.

Implementation details:

• Package imports work from the repository root. Data generation uses the checkout's oracle directory by default.
• Prompt metadata and tokenizers accompany intermediate checkpoints and the final adapter.
• Evaluation records selected metrics with explicit denominators and propagates subprocess failures.
• Data reuse requires a matching recorded recipe. Training protects existing final adapters and their configuration.
• Historical results remain available with explicit qualifications. Normalized scores are not calibrated correctness guarantees.

Validation:

• 20 offline unit tests passed, including synthetic scoring, padding, batching, rendering controls, metadata, and failure propagation.
• Infix and Lean pipeline dry runs passed.
• Module and script CLI smoke checks, compilation checks, and git diff --check passed.
• An offline GitHub Actions workflow runs the tests and dry runs.

Limitations:

• Real Gemma/Llama inference and a GPU training smoke test remain outstanding. Tests do not download base weights.
• The public pipeline uses the existing label-balanced sampler. Local shape-balanced v2/v3 work and shape diagnostics remain separate.
• The existing sampler permits repeated underlying pairs. Class-held-out evaluation requires at least one novel endpoint, not necessarily both.
• These judges predict relation tokens; they do not generate chain-of-thought text. This PR establishes no new causal interpretability result.
• Dataset reuse checks configuration, not the identity of the oracle files. Changed oracle artifacts require a fresh data directory.

Generated data and model weights are excluded. Keep this PR in draft until model smoke tests and review are complete.