Address PR #358 AI review round 5: align PR #354 CHANGELOG tail

P3 — the PR #354 [Unreleased] Fixed entry (line 21) said local-method
bit-identity SE remained blocked by the Rust-normalization and Python
cache-fallthrough divergences and was "tracked as a follow-up in
TODO.md." With the two TROP-local Fixed entries that this PR adds
(lines 22-27) closing exactly those divergences, the PR #354 tail
sentence is now internally inconsistent with the surrounding entries.
Rewritten to say the RNG half of finding #23 is closed here (bootstrap
contract), grid-search half was closed in PR #348, and the local-
method methodology half is closed by the two Fixed entries that
follow in the same release.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: igerber

CHANGELOG.md

@@ -18,7 +18,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - SyntheticDiD `variance_method="bootstrap"` now computes p-values from the analytical normal-theory formula using the bootstrap SE (matching R's `synthdid::vcov()` convention), rather than an empirical null-distribution formula that is not valid for bootstrap draws. `is_significant` and `significance_stars` are derived from `p_value` and will also change for bootstrap fits. Placebo and jackknife are unchanged. Point estimates are unaffected.
 - SyntheticDiD bootstrap SE formula applies the `sqrt((r-1)/r)` correction matching R's synthdid and the placebo SE formula.
 - SyntheticDiD bootstrap now retries degenerate resamples (all-control or all-treated, or non-finite `τ_b`) until exactly `n_bootstrap` valid replicates are accumulated, matching R's `synthdid::bootstrap_sample` and Arkhangelsky et al. (2021) Algorithm 2. Previously the Python path counted attempts (with degenerate draws silently dropped), producing fewer valid replicates than requested. A bounded-attempt guard (`20 × n_bootstrap`) prevents pathological-input hangs.
-- **TROP global bootstrap SE backend parity under fixed seed** — Rust and Python backends now produce bit-identical bootstrap SE under the same `seed`. Previously Rust's `bootstrap_trop_variance_global` seeded `rand_xoshiro::Xoshiro256PlusPlus` per replicate while Python's fallback consumed `numpy.random.default_rng` (PCG64), producing ~28% SE divergence on tiny panels under `seed=42`. Fixed by extracting a shared `stratified_bootstrap_indices` helper in `diff_diff/bootstrap_utils.py` that pre-generates per-replicate stratified sample indices via numpy on the Python side; both backends consume the same integer arrays through the PyO3 surface. Sampling law (stratified: controls then treated, with replacement) is unchanged. Closes silent-failures audit finding #23 (bootstrap half; grid-search half closed in PR #348). Local-method TROP also adopts the Python-canonical index contract for the RNG layer, but separately-discovered backend divergences (Rust normalizes weight-matrix outer product, Python `_compute_observation_weights` reads stale `_precomputed` cache) prevent local-method bit-identity SE; tracked as a follow-up in `TODO.md`.
+- **TROP global bootstrap SE backend parity under fixed seed** — Rust and Python backends now produce bit-identical bootstrap SE under the same `seed`. Previously Rust's `bootstrap_trop_variance_global` seeded `rand_xoshiro::Xoshiro256PlusPlus` per replicate while Python's fallback consumed `numpy.random.default_rng` (PCG64), producing ~28% SE divergence on tiny panels under `seed=42`. Fixed by extracting a shared `stratified_bootstrap_indices` helper in `diff_diff/bootstrap_utils.py` that pre-generates per-replicate stratified sample indices via numpy on the Python side; both backends consume the same integer arrays through the PyO3 surface. Sampling law (stratified: controls then treated, with replacement) is unchanged. Closes the bootstrap-RNG half of silent-failures audit finding #23 (grid-search half closed in PR #348; local-method methodology half closed by the two Fixed entries below). Local-method TROP also adopts the Python-canonical index contract for the RNG layer here.
 - **TROP local-method Rust weight-matrix no longer normalized** — `rust/src/trop.rs::compute_weight_matrix` no longer divides time-weights or unit-weights by their respective sums before the outer product. The paper's Equation 2/3 (Athey, Imbens, Qu, Viviano 2025) and REGISTRY.md Requirements checklist (line 2037: `[x] Unit weights: exp(-λ_unit × distance) (unnormalized, matching Eq. 2)`) both specify raw-exponential weights; Python's `_compute_observation_weights` was already REGISTRY-compliant. **User-visible effect**: Rust local-method ATT values may shift for any fit with `lambda_nn < infinity` — normalizing the weight-matrix inflated the effective nuclear-norm penalty relative to the data-fit term, changing the regularization trade-off. For `lambda_nn = infinity` (factor model disabled) outputs are unchanged because uniform weight scaling leaves the minimum-norm WLS argmin invariant. Rust LOOCV-selected lambdas may also shift on this boundary; both backends now converge on the same REGISTRY-compliant selection.
 - **TROP local-method Python `_compute_observation_weights` now uses the function-argument `Y, D` and treats all non-target units as donors** — two coupled changes that bring Python structurally in line with Rust and the paper's Eq. 2/3:
     1. Removed the `if self._precomputed is not None:` branch that silently substituted `self._precomputed["Y"]` / `["D"]` / `["time_dist_matrix"]` (original-panel cache populated during main fit) for the function-argument `Y, D`. Under bootstrap, `_fit_with_fixed_lambda` computes fresh `Y, D` from the resampled `boot_data` and passes them in; the helper was discarding those and recomputing unit distances from the original panel, so Python's local bootstrap resampled units but reused stale unit-distance weights. Rust's bootstrap was already correct (always consumed `y_boot, d_boot`).