fixes

• **run_es(..., rel_time = <col>)**: pass a pre-built event-time column (time relative to treatment) — e.g. the Time_to_Treatment from paneltools/fect's get.cohort() — and run_es() uses it verbatim as the event-study factor (i(rel_time, treatment, ref = baseline)) instead of computing relative time from time/timing. This normalises the previous time = event_time, timing = 0 idiom and, unlike that idiom, keeps never-treated controls (NA rel_time) in the estimation sample. Requires estimator = "twfe" and method = "classic"; the default calendar-difference behaviour is unchanged. See the new **Details** section of ?run_es for the three relative-time conventions (calendar difference, within-unit rank, pre-built) and when they diverge in sparse/unbalanced panels.

Diagnostics

• **CS** now names cohorts that are silently dropped because their base period g-1 is not observed (a real hazard in sparse panels, where a cohort can vanish from the by-cohort and event-study aggregates), and exposes them as attr(result, "dropped_cohorts") on both run_es() and calc_att().
• **SA** warns when the saturated design is degenerate (a large share of cohort×period interactions dropped as collinear, or implausibly large CATT standard errors), which is easy to miss with sparse/irregular time.
• **BJS** now reports attr(result, "n_unimputed") (and n_treated_obs): the number of treated observations whose counterfactual could not be imputed and were excluded from aggregation, on both run_es() and calc_att().

Internals — reduced fixest dependency

The long-term plan of trimming the fixest dependency (see CLAUDE.md)

begins in this release. Two internal engines now replace fixest::feols()

in most estimators; **all user-facing APIs, output columns, and numerics are

unchanged** — every change here is verified against the previous

fixest-based results to numerical agreement (tolerance = 1e-6, typically

much closer).

• **BJS Step 1** (fixed-effects nuisance regression) is now estimated by an internal alternating-projections solver (src/fe_solver.cpp) instead of fixest::feols(y ~ 0 | unit + time) + fixest::fixef(). It reproduces fixest's estimation sample exactly, including the iterative removal of singleton fixed-effect levels (dropping one singleton can create another).
• **SA, TWM, FLEX, and compute_contamination_weights()** now run their saturated cohort × period regressions through a new internal FE-OLS engine (R/fe_ols.R, src/demean_kway.cpp, src/fe_ols_kernels.cpp): k-way fixed-effects demeaning, keep-first collinearity dropping, and iid / heteroskedasticity-robust (HC1) / one-way cluster-robust standard errors that reproduce fixest's default small-sample corrections exactly (ssc(adj = TRUE, cluster.adj = TRUE), nested-fixed-effect accounting). Requests the internal engine cannot reproduce exactly — multi-way clustering, custom vcov_args, or fixed-effect specifications that are not plain columns — are transparently delegated to fixest, so every code path keeps returning identical results.
• The regressions are parallelised with OpenMP (falls back to sequential execution where OpenMP is unavailable) and are within a few percent of fixest's multi-threaded speed on saturated designs with hundreds of cohort × period columns. Thread count defaults to half the detected physical cores and can be set with options(fixes.nthreads = <n>).
• fixest remains the estimation engine for classic TWFE / method = "sunab" in run_es(), for run_did(), and as the fallback described above; it is still a required (Imports) dependency. parallel (base R) is now also imported, for core-count detection only.