Add PIR (Physics Intermediate Representation) method

[Qazi-pk]

Method: PIR (Physics Intermediate Representation)

A torch-free symbolic regression method combining dimensional analysis,
RANSAC consensus filtering, residual refinement, and sparse coefficient
selection. Sklearn-compatible.

• Engine source: https://github.com/Qazi-pk/physics-engine (MIT, v0.1.0, commit 736a89c)
• License: MIT
• Python: 3.10+
• Dependencies: numpy, sympy, scikit-learn, scipy (no torch)

Files added

• experiment/methods/PIRRegressor.py — exports est, hyper_params,
  complexity(est), model(est, X)
• experiment/methods/src/PIR_install.sh — pip-installs physics-engine@v0.1.0
  from the public source repo (no source code vendored in this PR)

Configuration

Blind-sweep settings (SEED=0):
enforce_dimensions=False, allowed_powers=[1, 2],
include_pairwise_products=True, use_ransac=True, use_residual=True,
use_sparse=True, use_ot_loss=False, add_physics_features=False.
Other parameters at defaults.

Honest scope

Blind Tier A (Feynman, SRBench-compatible protocol, 5 seeds):
7/44 solved (≈7.6 mean) under the configuration above. An earlier
27.3% figure in the project's own notes was formula-peeking and is not
the blind result.

The current classical engine is architecturally limited to ≤ 2-variable
monomial structures (pairwise structure detector); extension to ≥ 3-variable
laws is future work and not part of this PR.

Open question for reviewers

model(est, X) returns a sympy-parseable string using the column names of
the input pd.DataFrame X. If SRBench's symbolic-equivalence check
expects a specific naming convention (e.g. x_0..x_m vs Feynman's q1,
Ef, …), please flag — happy to adjust the variable mapping in
model() to match.

Checklist

• Targets dev branch
• Sklearn-compatible API (fit, predict, random_state)
• No source code vendored; install.sh pulls from stable tag
• model(est, X) returns a sympy-compatible string
• MIT licensed

[Qazi-pk]

CI is failing in Set up job with actions/cache@v2 deprecated — this appears to be an upstream workflow issue, not specific to this PR (no project files are reached before the failure).

Happy to wait while it's addressed, or to open a small follow-up PR bumping the cache action version if that would help.