[igerber]

Summary

• 4.0 cut, PR 3b of three (3a test(v4): PR 3a of the 4.0 cut - migrate the suites to the post-fit aggregation containers #863 migrated the suites; 3c carries the ImputationDiD / TwoStageDiD / ContinuousDiD removals plus the bootstrapped-fit relay). This is the mechanical half: the fit-time aggregate= / balance_e= keywords are removed where the post-fit route already exists - CallawaySantAnna (M-020/M-117), EfficientDiD (M-023/M-120), StackedDiD (M-024), ChaisemartinDHaultfoeuille (M-026), HeterogeneousAdoptionDiD (M-027) and did_had_pretest_workflow (M-139). The eight ledger rows flip to removed; passing the keywords is now Python's native TypeError.
• CallawaySantAnna and EfficientDiD compute no event-study/group surface at fit: the legacy fields (and CS's cband_crit_value / event_study_vcov / event_study_vcov_index / event_study_df) are None after every fit, and the post-fit replay is the only bootstrap route (unchanged). StackedDiD's fit-time group/all ValueErrors and the HAD _VALID_AGGREGATES sets go with the keywords; _infer_aggregate_mode is HAD's sole selector. Positional slots after the removed parameters shift left (documented in the migration guide).
• Runtime messages (honest_did, pretrends, plot_event_study, _reporting_helpers, diagnostic_report), the four llms*.txt guides, REGISTRY / REPORTING notes, the migration guide, API and performance docs stop steering users to the removed route on these five estimators. Wording about ImputationDiD / TwoStageDiD / ContinuousDiD stays true at this PR (their keyword is still deprecated, not removed). docs/v4-design.md section 6 gains the 3b amendment and section 9 marks sub-item 3b shipped; ledger row M-093 records that CS/EDiD populate the legacy fields on no fit.
• tests/test_aggregate_contract.py: the five Test*Shim classes, the fit-time fixtures (fit_time, efficient_fit_time, _boot_fit_time, _efficient_boot_fit_time) and the inertness comparisons are retired for Test*Removal pins (TypeError, signature, positional-slot binding); the committed oracle fixture stays the numerical reference. tests/_capture_v4_aggregate_oracles.py is marked historical (it drives the removed route; runnable only at the recorded captured_at_head). warn_deprecated_kwarg loses its last library caller and is retained per spec section 8 rule 11 with a direct unit pin.

Methodology references (required if estimator / math changes)

• Method name(s): N/A - no estimator mathematics changes (parameter removals, message and documentation sweep; every retained numerical surface is pinned against the committed per-backend oracles captured on the pre-removal tree)
• Paper / source link(s): N/A
• Any intentional deviations from the source (and why): None

Validation

• Tests added/updated: tests/test_aggregate_contract.py (437 tests, numpy and rust backends), tests/test_staggered.py, tests/test_efficient_did.py, tests/test_stacked_did.py, tests/test_chaisemartin_dhaultfoeuille.py, tests/test_had.py, tests/test_had_pretests.py, tests/test_had_dual_knob_deprecation.py, tests/test_v4_rename_mechanical.py, tests/test_guides.py, tests/test_naming_guard.py, tests/test_event_study_surface.py, comment-only wording in nine CS/EDiD consumer suites. The touched and consumer suites (35 files, 4,132 tests) pass; test_v4_matrix sees the eight rows flip (due-at-4.0 non-terminal rows 24 -> 16). mypy diff_diff clean; Sphinx -W build clean; tutorial 02 executes.
• Backtest / simulation / notebook evidence (if applicable): N/A

Security / privacy

• Confirm no secrets/PII in this PR: Yes

Changelog

• changelog.d/ fragment added (or N/A - no user-visible change): Yes - changelog.d/20261021-v4-pr3b-aggregate-removals-mechanical.md (### Removed, ### Breaking Changes, ### Changed, ### Internal)

Follow-ups

• PR 3c: the ImputationDiD / TwoStageDiD / ContinuousDiD removals (M-021/M-022/M-025/M-118/M-119) plus the bootstrapped-fit relay (spec section 6 amendment), prepared on top of this branch and opened once it merges.

[github-actions]

Overall assessment: ✅ Looks good

No unmitigated P0 or P1 findings identified.

Executive summary

• The five estimator removals and HAD workflow removal consistently enforce the intended API.
• No changes to estimands, weighting, identification checks, or variance formulas were found.
• Bootstrap replay remains available through post-fit aggregation.
• Breaking changes have Registry notes, migration guidance, and a changelog fragment.
• Numerical tests could not run: the available Python lacks pytest, NumPy, and pandas.

Methodology

Reviewed Callaway–Sant’Anna, EfficientDiD, StackedDiD, dCDH, and HAD/pretests against Registry equations, in-code references, and available on-file paper reviews.

• P3 — Documented API transition. Removing fit-time aggregation and leaving CS/EfficientDiD legacy fields empty matches the explicit Registry notes (docs/methodology/REGISTRY.md:L1284, L1873, L2468, L5022).
  • Impact: Callers must use post-fit containers; positional parameters shift as documented.
  • Concrete fix: None required. These are documented behavior changes, not methodology defects.

No undocumented methodology deviation or newly incorrect inference was identified.

Code Quality

No findings. Static caller checks found no remaining library or benchmark fit(aggregate=...)/fit(balance_e=...) calls. No new inline-inference anti-patterns or incomplete parameter propagation were found.

Performance

No findings. Bootstrap replay retains its documented per-call cost; this PR does not introduce a new replay implementation.

Maintainability

No findings. Public signatures, removal tests, ledger statuses, and runtime guidance agree.

Tech Debt

No new untracked limitation identified. Historical oracle capture instructions explain how to reproduce the committed reference data after removal.

Security

No security issues or accidental secrets identified in the changes.

Documentation/Tests

No actionable findings. The changelog fragment covers the user-visible removals, and regression coverage retains numerical oracles plus keyword-rejection and positional-binding checks.

Validation completed: syntax checks for 18 modified modules and seven dependency-free guide assertions passed. Full pytest execution and numerical oracle verification remain unverified in this environment.