Extend cd to mlsynth's feature set (PR B of 4)

[cmg777]

Summary

Stacked on #2. Once that merges, this PR retargets to main. It adds the Cao & Dowd features that mlsynth's SPILLSYNTH(method="cd") has and the merged cd model lacked. Each feature follows the paper's v3 definitions and is cross-checked against mlsynth.

Feature	v3 reference	API
Homogeneous spillovers	Example 2 (Sec. 3.4)	spillover_structure="homogeneous" with affected_units
Exponential distance decay	Example 3 (Sec. 7.1)	spillover_structure="distance_decay" with unit_distances
Several treated units, common start	Sec. S.1.2	detected from the treat column; att_by_unit
Efficient (Ω⁻¹-weighted) estimator	Sec. S.1.1	weighting="efficient", which fills result.cd.efficient
Pure-donor sensitivity	Sec. 5.2.2	result.cd.sensitivity; plot kind "sensitivity"
Structure selection by κ_A	Secs. 5.1.2 and S.1.3	result.select_structure_by_kappa(candidates)

Two choices worth a look:

• Selection statistic. By default, structure selection uses the paper's multi-period κ_A (S.1.3): the norm of the summed residual vectors, divided by √T1. mlsynth instead averages the per-period norms. That variant is available as statistic="mean_norm", and the parity test uses it.
• Efficient estimator. It is reported alongside the headline identity-weighted fit, point estimates only, as in mlsynth. It warns when T0 ≤ number of units, because Ω̂ is then singular without its ridge. That is the case for Proposition 99 (T0 = 19).

Validation

• mlsynth fixtures. tools/freeze_mlsynth_cd_reference.py runs mlsynth 1.0.0 (cvxpy/CLARABEL) in a separate environment and writes benchmarks/reference/mlsynth_cd/reference.json. The README there records provenance and SHA-256s.
• Parity test. test_cd_mlsynth_parity.py checks six cases plus structure selection. Coverage:
  • SCM weights within 1e-4, the difference between the two solvers;
  • effects, CIs, joint spillover test, κ_A, sensitivity weights and the efficient fit;
  • P values exactly (within one tie).
• Live comparison. It also passes against a live mlsynth install, in a scratch environment.
• Invariants. test_cd_features.py checks:
  • equal distances give the homogeneous structure;
  • per-unit and homogeneous agree with one affected unit;
  • a custom A equal to a built A gives the same results, and column rescaling only rescales γ;
  • W = I recovers the unweighted estimator;
  • the choice of headline unit doesn't change the effects;
  • staggered adoption is rejected, and sar still rejects two treated units.
• Authors' code. The homogeneous κ_A matches the authors' code.
• Full suite. make lint typecheck test: 306 passed, 3 skipped. Slow tests pass.

🤖 Generated with Claude Code