…on; summary alpha

Addresses unbalanced-panel correctness:
- Reweight equal-weighting denominators now come from the realized post-drop
  sample (was the pre-drop panel), preserving the Callaway-Sant'Anna
  equivalence on unbalanced panels.
- Regression-adjustment path drops treated observations at event times with no
  clean control (counterfactual unidentified) with a UserWarning, and guards the
  SE to stay NaN-consistent with a NaN point estimate (the prior groupby-sum
  vectorization silently turned all-NaN scores into 0.0).
- Validate pre_window / post_window as non-negative integers.
- LPDiDResults.summary() no longer accepts an alpha override (it would relabel
  the CI level without recomputing the displayed intervals).
- Docs: clarify no_composition fixes the post-treatment composition; reword the
  api/lpdid.rst covariate-path wording (RA is preferred, not auto-default).
- Tests: unbalanced reweight==CS, RA NaN-consistency, fixed post composition,
  window validation.

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

…ation

- PMD paths require the ACTIVE baseline column (the premean column) rather than
  always the exact t-1 outcome, so an unbalanced-panel observation with a valid
  PMD baseline but a missing t-1 outcome is no longer silently dropped.
- Reject empty pooled pre windows (pre_window < 2 with pooled output requested)
  with a clear ValueError; exclude the h=-1 reference horizon from the supported
  pooled-pre window (it would inject mechanical zero long differences).
- Tests for PMD missing-t-1 retention, empty pooled pre window, and -1 rejection.

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

…ficient_action

- no_composition now requires every post target (h=0..H) observed, not just the
  maximum horizon, so post-horizon samples stay fixed under non-monotone
  unbalanced missingness; the fixed-composition mask is applied only to post
  event horizons and the pooled-post window (pre-placebos use available data).
- Propagate the public rank_deficient_action parameter into LPDiDResults and
  to_dict().
- Docs: qualify the variance-weighted "stacked" equivalence as Cengiz et al.
  (2019)-style, not diff-diff's Wing-et-al StackedDiD.
- Tests: non-monotone no_composition fixed post sample, rank_deficient_action
  propagation.

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

The regression-adjustment covariate path uses an influence-function cluster
variance (ImputationDiD/BJS family), not an OLS CR1 sandwich. Record
vcov_type="if_cluster" for that path (reweight + covariates/lags/absorb) and
render an accurate summary label ("Influence-function cluster-robust"), so the
results metadata and summary no longer mislabel it as hc1 / CR1.

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

…rols warn

- pmd="max" premean baseline now divides by the count of NON-MISSING prior
  outcomes (not prior rows), so a present-but-NaN pretreatment outcome no longer
  deflates the baseline and silently biases the estimate (cumsum already skips
  NaN, so only the denominator was wrong).
- The direct-inclusion homogeneity warning (reweight=False) now also fires for
  ylags / dylags, which are direct-included controls subject to the same
  non-negative-weight caveat (online Appendix B.2.2); warning reworded to
  "covariate-style controls".
- Tests: pmd="max" NaN-history exclusion, ylags/dylags warning.

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

… lag metadata

- no_composition now requires the active baseline and every post-treatment
  target OUTCOME to be non-missing (value-based via reindex), so the fixed
  post composition holds under any missingness encoding -- absent rows OR
  present-but-NaN outcomes -- not merely row existence.
- The pooled-window identification pre-check uses base identification
  (apply_no_composition=False); pooled estimation applies its own pooled-window
  composition mask, so a narrower post_pooled with missing far-horizon data is
  no longer spuriously rejected as unidentified.
- Record ylags / dylags on LPDiDResults and to_dict() for auditability.
- Reject bool pre_pooled / post_pooled (bool is an int in Python).
- Tests: present-but-NaN no_composition, bool rejection, lag metadata.

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

…e dead code

- Validate alpha in (0, 1) and require a numeric `time` column (clear error
  instead of a cryptic downstream failure on irregular/datetime labels).
- Results-level n_clusters now reports the realized cluster count of the
  pooled-post headline row (the summary "G"), not the full-panel count; per-row
  realized counts remain in the event_study / pooled tables.
- Remove the now-unused _outcome_available_mask (superseded by the value-based
  availability check in _common_clean_sample_indicator).
- Document the numeric-time requirement in api/lpdid.rst.
- Tests for alpha / time validation and headline n_clusters.

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

…aN clusters

- Validate the global time grid is integer-spaced by 1 (not merely numeric):
  irregular grids (e.g. 2000, 2002, ...) raise rather than silently producing
  empty t+h horizons / inconsistent horizon meanings.
- Reject str covariates / absorb (would iterate character-by-character).
- Reject missing values in the effective cluster column (they would silently
  drop from the RA cluster variance via groupby, biasing SEs with no warning).
- Tests for all three.

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

_prepare_panel built outcome lags, first differences, and integer-pmd premean
baselines (plus treatment-entry detection) with row-order ops (shift/diff/
rolling), which equate "previous observed row" with calendar t-1 -- correct
only on a regular per-unit grid. On a unit with an interior time gap (observed
t=0,2 missing t=1, possible in an unbalanced panel even when the global grid is
consecutive) they silently used the wrong period.

Fix: reindex each unit to its complete interior calendar grid, compute the
features on the grid (every row-order op now indexes true calendar time), then
restrict back to the observed rows -- a lag/difference spanning a gap is NaN
and the observation fails closed; no synthetic NaN-cluster gap row reaches a
regression or the reweight denominators. A gap-free panel skips the reindex
(early-out) and is bit-identical. Absorbing/cluster validation runs on observed
rows before reindex; treatment is absorbing-filled on the grid (exact on
observed rows). Also require integer-valued time labels (the spacing check
admitted fractional 0.5, 1.5).

Documented in REGISTRY (interior-gap handling + "entry = first observed
treated" convention). 7 new interior-gap tests; existing 56 unchanged.

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

_pmd_all_baseline subtracted panel[outcome] (the base row's own y_t), so a base
row with a missing current outcome got a NaN premean baseline and was silently
dropped at every horizon -- even though PMD's premean uses only PRIOR outcomes
and the long difference is y_{t+h} - premean (y_t is never used). A treated
entry whose entry-period outcome is missing thus lost all treatment variation,
yielding NaN coefficients. Build the numerator from fillna(0) cumulative sums
(the strictly-prior non-missing sum), independent of y_t; bit-identical when no
outcome is NaN. Regression test: pmd="max" retains the h>0 treated obs when the
entry-period outcome is missing.

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

- RA path (reweight=True with absorb): a treated observation whose absorbed
  level has no clean-control support has an all-zero control dummy, so its
  counterfactual would extrapolate through an unidentified coefficient. Drop
  those treated obs with a warning (mirrors the existing event-time
  identification check), never impute off a non-identified fit.
- Document the pooled pre/post estimand in REGISTRY (Note 6): the unit-equal-
  weighted average of each unit-event-time's mean long difference on the
  fixed-composition sample (every pooled target observed); equals the mean of
  the per-horizon event-study coefficients on a balanced panel, and differs
  from the authors' horizon-stacked pooled regression under cross-horizon
  composition changes -- exact parity reconciled in PR-B2.
- Tests: RA unsupported-absorbed-level drop; pooled-post == mean event-study.

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

…rol label

- Reject covariates/absorb names colliding with LPDiD working columns (names
  starting with "_" or "horizon") before panel construction, so a user column
  cannot silently overwrite an internal column.
- Move the covariates/absorb string-vs-list check before the required-columns
  build, so absorb="region" raises the precise "must be a list" error rather
  than an iterate-characters missing-column error.
- Relabel the summary's "Control units" -> "Never-treated units" (the count is
  never-treated units; clean controls also include not-yet-treated cohorts).
- Check off the B1 pure-Python test checklist row in REGISTRY.

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

…ol_units

- to_dict() now includes covariates and absorb (the result stores them and
  summary() displays them; serialized results were dropping the adjustment
  specification).
- Document that LPDiDResults.n_control_units counts never-treated units only
  (the library-wide field convention; the realized clean-control pool also
  includes not-yet-treated cohorts, whose per-horizon counts are in the table
  columns). summary() already labels it "Never-treated units".
- Test docstring: external R-parity "will live in" test_methodology_lpdid.py
  (added in PR-B2), not present in this PR.

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

…LS path

The RA path drops treated observations at event times with no clean control,
but the default (reweight=False) path only checked global treated/control
presence. A treated event time with no clean control makes its time fixed
effect collinear with the treatment indicator; the rank handler could drop that
time dummy and identify the effect off invalid cross-event-time comparisons,
yielding a spurious finite estimate (e.g. the last-treated cohort under
control_group="clean" with no never-treated units). Mirror the RA event-time
identification check in _estimate_sample (covers both the default event-study
and the default pooled path): drop unsupported treated obs with a warning, NaN
if none remain. A heterogeneous-effect regression test proves the unidentified
cohort no longer contaminates the estimate.

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

…n docs

- summary() now lists nonzero ylags/dylags on the controls line (covariates and
  absorb were already shown).
- llms-full.txt: clarify SEs are unit-cluster CR1 on the default/weighted path
  but the influence-function cluster variance on the RA covariate path.
- llms-full.txt + api/lpdid.rst: no_composition fixes the POST-treatment
  composition across POST horizons (pre-placebos may vary), matching the
  implementation.

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

The RA-vs-default finite-sample-factor note said the asymmetry "is validated
against the reference R packages" while the same section marks R parity as
pending (PR-B2). Reword to "is documented here and will be validated ... in the
R-parity follow-up" so the status is not overclaimed.

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

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