Address PR #456 R5 review (1 P1 + 1 P2 + 1 P3)

P1: reject horizon_max=0 under event_study=True. The previous docs said
H=0 was a "well-defined but semantically distinct" design, but every
event_study=True + horizon_max=0 + anticipation=0 fit hit the ref_period
guard at -1 and raised. Resolution: lock the contract by rejecting H=0
explicitly with a remediation message ("use event_study=False for the
aggregate Wave B spec; event-study mode requires horizon_max>=1 or
horizon_max=None"). Updated REGISTRY + CHANGELOG to match. Added
test_horizon_max_zero_with_event_study_raises regression.

P2: plot_event_study now honors SpilloverDiDResults.reference_period.
Wave C's rectangular event_study_effects emits multiple empty horizons
(n_obs == 0 on dropped post-direct cells + the reference row); the legacy
"first n_obs==0 row" reference detection could pick a non-reference
empty horizon as the reference. Fix in _extract_plot_data: prefer
results.reference_period when present (truthy attribute), fall back to
the legacy n_obs==0 heuristic otherwise. Backward-compatible for
estimators without the attribute (CallawaySantAnna, SunAbraham, etc.).
Regression test on a Wave C fit with horizon_max=4 (oversized → multiple
empty horizons) asserts the inferred reference is -1 not the first empty
horizon.

P3: soften "Wave B bit-identical" claim. CHANGELOG previously said
"reproduces Wave B SEs bit-identically (verified by ...)" implying a
pre-Wave-C checkout artifact; the goldens were actually captured on
the current (Wave C) event_study=False path. Updated to: "the aggregate
stage-2 design construction, fit, and extraction logic on this path
are byte-identical to Wave B; the test pins goldens captured on the
unchanged aggregate path so any future drift fails the regression."
Same softening in the test class docstring.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Author: igerber