…ADH 2015 / Abadie 2021)

Completes the ADH-2015 / Abadie-2021 predictor-importance V-selection menu with two new
`v_method` values, each threaded through the in-space / leave-one-out / in-time placebo
refits so a diagnostic uses the same estimator as the headline fit.

`v_method="cv"` — out-of-sample cross-validation (ADH 2015 §; Abadie 2021 Eq. 9). The
pre-period is split at `v_cv_t0` (new constructor param; default `len(pre)//2`) into a
training and a validation window. Each predictor spec is RE-AGGREGATED over each window
(its op recomputed over only that window's periods — a separate `dataprep` per window,
standardized per window), so the V-search is genuinely out-of-sample for every predictor
type: V is selected to minimize the validation-window outcome MSPE of the training-window
fit, then the final weights are re-estimated on the validation-window predictors (step 4).
The same V* drives both fits with no zeroed coordinate, so `v_weights` reproduce
`donor_weights` and `predictor_balance` is reported on the validation-window basis.
`mspe_v` reports the held-out validation MSPE. Abadie 2021 fn.7 non-uniqueness is handled
by a deterministic, convergence-aware flat-MSPE tie-break (prefer the densest V; never let
a non-converged candidate displace a converged incumbent).

`v_method="inverse_variance"` — closed-form `v_h = 1/Var(X_h)` (Abadie 2021 §3.2(a)),
variance over donors+treated on the unstandardized predictors, applied to the RAW
predictors (the `standardize` pre-scaling is intentionally bypassed — inverse-variance
weighting IS the unit-variance rescaling). Exact for every positive variance (no flooring);
zero-variance rows get 0 weight; an all-zero / overflow panel falls back to uniform + warn.

Fail-closed identification gates (cv): every predictor must SPAN both windows (re-aggregation
needs it measurable on each — default single-period lags are rejected), and each window must
have cross-DONOR variation (donor-indistinguishable windows leave X0·W constant in W → the
weight solve is unidentified). Violations raise on the headline fit; in-space placebos drop
the affected refit; in-time-truncated dates are marked `infeasible`. Single-donor fits force
w=[1] so V is unidentified → uniform `v_weights` + `mspe_v=None` for all methods (documented).

Validation: R `Synth` has no built-in CV, so cv is anchored by deterministic equivalence to
the R-anchored `custom_v` path on the per-window re-aggregated predictors (both the step-3
criterion and step-4 weights) + cv self-consistency (in_time_placebo == fresh backdated fit);
inverse-variance is bit-for-bit vs `custom_v=1/Var(X)`. New tests cover config validation,
exact inverse-variance, the spanning / donor-variation / convergence fail-closed gates,
single-donor degeneracy, and placebo/LOO/in-time propagation for both methods.

Docs: REGISTRY §SyntheticControl Notes (per-window re-aggregation, fully-spanning + donor
gates, tie-break, inverse-variance, single-donor), checklist tick, `docs/api`, the LLM
guides, README, CHANGELOG. The remaining ADH-2015 tail (`W^reg` extrapolation, sparse-SC)
and an in-space/LOO machine-readable cv-infeasible reason-code stay tracked in TODO.md.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>