Address PR #393 R1 P1: multi-baseline trends_linear UserWarning

CI reviewer flagged that by_path + trends_linear was ungated on all
panels but only warned on F_g<3 exclusions, while the analogous
by_path + controls path warns on multi-baseline switcher panels
(D_{g,1} multiplicity) where R's per-path full-pipeline call diverges
from Python's global-then-disaggregate architecture. The by_path
docstring already acknowledged the "same multi-baseline divergence
pattern as controls" (line 461-462), but the runtime warning was
missing — a docstring-vs-implementation gap.

Fix: mirror the controls warning at chaisemartin_dhaultfoeuille.py
:1565-1584 inside the trends_linear block (right after the F_g<3
warning). Same predicate (switcher mask on first_switch_idx_arr,
unique baseline check). Different message naming the trends_linear
mechanism (full-pipeline including first-differencing per path
subset).

REGISTRY note updated to spell out the warning explicitly.

Three regression tests added in TestByPathTrendsLinear, mirroring
TestByPathControls:
- test_multi_baseline_panel_emits_r_deviation_warning: panel with
  joiners and leavers must emit the warning
- test_single_baseline_panel_does_not_emit_r_deviation_warning:
  single-baseline panel must not warn
- test_single_baseline_heterogeneous_F_g_does_not_warn: pins the
  precise predicate (baseline multiplicity, NOT F_g multiplicity)

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

Author: igerber

diff_diff/chaisemartin_dhaultfoeuille.py

@@ -1653,6 +1653,41 @@ def fit(
                     UserWarning,
                     stacklevel=2,
                 )
+            # Multi-baseline switcher panel detection (`by_path + trends_linear`):
+            # mirror the analogous warning fired by `by_path + controls` at
+            # `:1565-1584`. Python first-differences once globally before
+            # path enumeration; R `did_multiplegt_dyn(..., by_path, trends_lin)`
+            # re-runs the full pipeline (including first-differencing) per
+            # path's restricted subsample. On single-baseline switcher
+            # panels the two architectures coincide; on multi-baseline
+            # switcher panels (some switchers have `D_{g,1}=0`, others
+            # `D_{g,1}=1`) per-path point estimates can diverge — warn the
+            # user explicitly so they don't silently consume estimates
+            # that disagree with R. The check filters to switcher groups
+            # only (never-switchers / always-treated controls don't
+            # contribute to switcher baseline multiplicity).
+            if self.by_path is not None:
+                _switcher_mask_tl = first_switch_idx_arr >= 0
+                if _switcher_mask_tl.any():
+                    _switcher_baselines_tl = baselines[_switcher_mask_tl]
+                    if np.unique(_switcher_baselines_tl).size > 1:
+                        warnings.warn(
+                            "by_path + trends_linear: switcher baselines "
+                            "D_{g,1} take multiple values in this panel. "
+                            "Python first-differences once on the full "
+                            "panel before path enumeration; R "
+                            "`did_multiplegt_dyn(..., by_path, trends_lin)` "
+                            "re-runs the full pipeline (including "
+                            "first-differencing) on each path's restricted "
+                            "subsample, so per-path point estimates can "
+                            "diverge between Python and R on this panel. "
+                            "See `docs/methodology/REGISTRY.md` "
+                            "(`Note (Phase 3 by_path ...)` -> Per-path "
+                            "linear-trends DID^{fd}) for the full "
+                            "deviation contract.",
+                            UserWarning,
+                            stacklevel=2,
+                        )
             N_mat_orig = N_mat.copy()
             Y_mat, N_mat = _compute_first_differenced_matrix(Y_mat, N_mat)