Narrow negative-D claim to tested scope; track baseline gap in TODO

R1 review on the prior commit flagged that the prose said
"negative-integer baseline support is regression-tested in Python only"
but the visible regression only exercises paths with negative D values
in non-baseline positions (e.g. `(0, -1, -1, -1)`), not negative-
baseline paths like `(-1, 0, 0, 0)` - which is the exact regime that
would trigger R's `substr(path, 1, 1)` parser bug.

Narrow the doc claim to match what's actually tested ("negative-D
treatment-state support in non-baseline positions") and add a TODO
entry to track the deferred negative-baseline regression so the gap
is explicit and the doc claim is accurate.

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

Author: igerber

CHANGELOG.md

@@ -12,7 +12,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - **Tutorial 21: HAD Pre-test Workflow** (`docs/tutorials/21_had_pretest_workflow.ipynb`) — composite pre-test walkthrough for `HeterogeneousAdoptionDiD` building on Tutorial 20's brand-campaign framing. Uses a 60-DMA × 8-week panel close in shape to T20's but with the dose distribution drawn from `Uniform[$0.01K, $50K]` (vs T20's `[$5K, $50K]`); the true support is strictly positive but very near zero, chosen so the QUG step in `did_had_pretest_workflow` fails-to-reject `H0: d_lower = 0` in this finite sample and the verdict text fires the load-bearing "Assumption 7 deferred" pivot for the upgrade-arc narrative. (HAD's `design="auto"` selector — a separate min/median heuristic at `had.py::_detect_design`, NOT the QUG p-value — independently lands on the `continuous_at_zero` identification path with target `WAS` on this panel because `d.min() < 0.01 * median(|d|)`. The QUG test and the design selector are independent rules that point to the same identification path here.) Walks through three surfaces: (a) `did_had_pretest_workflow(aggregate="overall")` on a two-period collapse, where the verdict explicitly flags Step 2 (Assumption 7 pre-trends) as not run because a single pre-period structurally cannot support a pre-trends test, and the structural fields `pretrends_joint` / `homogeneity_joint` are both `None`; (b) `did_had_pretest_workflow(aggregate="event_study")` on the full multi-period panel, where the verdict reads "TWFE admissible under Section 4 assumptions" because all three testable diagnostics (QUG + joint pre-trends Stute over 3 horizons + joint homogeneity Stute over 4 horizons) fail-to-reject — non-rejection evidence under finite-sample power and test specification, not proof that the identifying assumptions hold; and (c) a side panel exercising both `yatchew_hr_test` null modes — `null="linearity"` (default, paper Theorem 7) vs `null="mean_independence"` (Phase 4 R-parity with R `YatchewTest::yatchew_test(order=0)`) — on the within-pre-period first-difference paired with post-period dose, illustrating the stricter null's larger residual variance (`sigma2_lin` 7.01 vs 6.53) and smaller p-value (0.29 vs 0.49). Companion drift-test file `tests/test_t21_had_pretest_workflow_drift.py` (16 tests pinning panel composition, both verdict pivots, structural anchors on both paths, deterministic QUG / Yatchew statistics, bootstrap p-value tolerance bands per `feedback_bootstrap_drift_tests_need_backend_tolerance`, and `HAD(design="auto")` resolution to `continuous_at_zero` on this panel). T20's "Composite pretest workflow" Extensions bullet updated with a forward-pointer to T21. T22 weighted/survey HAD tutorial remains queued as a separate notebook PR.
 - **`ChaisemartinDHaultfoeuille.by_path` and `paths_of_interest` now compose with `survey_design`** for analytical Binder TSL SE and replicate-weight bootstrap variance. The `NotImplementedError` gate at `chaisemartin_dhaultfoeuille.py:1233-1239` is replaced by a per-path multiplier-bootstrap-only gate (`survey_design + n_bootstrap > 0` under by_path / paths_of_interest still raises, since the survey-aware perturbation pivot for path-restricted IFs is methodologically underived). Per-path SE routes through the existing `_survey_se_from_group_if` cell-period allocator: the per-period IF (`U_pp_l_path`) is built with non-path switcher-side contributions skipped (control contributions are unchanged, matching the joiners/leavers IF convention; preserves the row-sum identity `U_pp.sum(axis=1) == U`), cohort-recentered via `_cohort_recenter_per_period`, then expanded to observations as `psi_i = U_pp[g_i, t_i] · (w_i / W_{g_i, t_i})`. Replicate-weight designs unconditionally use the cell allocator (Class A contract from PR #323). New `_refresh_path_inference` helper post-call refreshes `safe_inference` on every populated entry across `multi_horizon_inference`, `placebo_horizon_inference`, `path_effects`, and `path_placebos` so all four surfaces use the same final `df_survey` after per-path replicate fits append `n_valid` to the shared accumulator. Path-enumeration ranking under `survey_design` remains unweighted (group-cardinality, not population-weight mass). Lonely-PSU policy stays sample-wide, not per-path. Telescope invariant: on a single-path panel, per-path SE matches the global non-by_path survey SE bit-exactly. **No R parity** — R `did_multiplegt_dyn` does not support survey weighting; this is a Python-only methodology extension. The global non-by_path TSL multiplier-bootstrap path is unaffected (anti-regression test `tests/test_chaisemartin_dhaultfoeuille.py::TestByPathSurveyDesignAnalytical::test_global_survey_plus_n_bootstrap_still_works` locks the per-path-only scope of the new gate). Cross-surface invariants regression-tested at `TestByPathSurveyDesignAnalytical` (~17 tests across gate / dispatch / analytical SE / replicate-weight SE / per-path placebos / `trends_linear` composition / unobserved-path warnings / final-df refresh regressions) and `TestByPathSurveyDesignTelescope`. See `docs/methodology/REGISTRY.md` §`ChaisemartinDHaultfoeuille` `Note (Phase 3 by_path ...)` → "Per-path survey-design SE" for the full contract.
 - **Inference-field aliases on staggered result classes** for adapter / external-consumer compatibility. Read-only `@property` aliases expose the flat `att` / `se` / `conf_int` / `p_value` / `t_stat` names (matching `DiDResults` / `TROPResults` / `SyntheticDiDResults` / `HeterogeneousAdoptionDiDResults`) on every result class that previously only carried prefixed canonical fields: `CallawaySantAnnaResults`, `StackedDiDResults`, `EfficientDiDResults`, `ChaisemartinDHaultfoeuilleResults`, `StaggeredTripleDiffResults`, `WooldridgeDiDResults`, `SunAbrahamResults`, `ImputationDiDResults`, `TwoStageDiDResults` (mapping to `overall_*`); `ContinuousDiDResults` (mapping to `overall_att_*`, ATT-side as the headline, ACRT-side accessible unchanged via `overall_acrt_*`); `MultiPeriodDiDResults` (mapping to `avg_*`). `ContinuousDiDResults` additionally exposes `overall_se` / `overall_conf_int` / `overall_p_value` / `overall_t_stat` aliases for naming consistency with the rest of the staggered family. Aliases are pure read-throughs over the canonical fields — no recomputation, no behavior change — so the `safe_inference()` joint-NaN contract (per CLAUDE.md "Inference computation") is inherited automatically (NaN canonical → NaN alias, locked at `tests/test_result_aliases.py::test_pattern_b_aliases_propagate_nan`). The native `overall_*` / `overall_att_*` / `avg_*` fields remain canonical for documentation and computation. Motivated by the `balance.interop.diff_diff.as_balance_diagnostic()` adapter (`facebookresearch/balance` PR #465) which calls `getattr(res, "se", None)` / `getattr(res, "conf_int", None)` without a fallback chain — pre-alias, every staggered result class returned `None` on those keys, silently dropping `se` and `conf_int` from the adapter's diagnostic dict. 23 alias-mechanic + balance-adapter regression tests at `tests/test_result_aliases.py`. Patch-level (additive on stable surfaces).
-- **`ChaisemartinDHaultfoeuille.by_path` + non-binary integer treatment** — `by_path=k` now accepts integer-coded discrete treatment (D in Z, e.g. ordinal `{0, 1, 2}`); path tuples become integer-state tuples like `(0, 2, 2, 2)`. The previous `NotImplementedError` gate at `chaisemartin_dhaultfoeuille.py:1870` is replaced by a `ValueError` for continuous D (e.g. `D=1.5`) at fit-time per the no-silent-failures contract — the existing `int(round(float(v)))` cast in `_enumerate_treatment_paths` is now defensive (no-op for integer-coded D). Validated against R `did_multiplegt_dyn(..., by_path)` for D in `{0, 1, 2}` via the new `multi_path_reversible_by_path_non_binary` golden-value scenario (78 switchers, 3 paths, single-baseline custom DGP, F_g >= 4): per-path point estimates match R bit-exactly (rtol ~1e-9 on event horizons; rtol+atol envelope for placebo near-zero values), per-path SE inherits the documented cross-path cohort-sharing deviation (~5% rtol observed; SE_RTOL=0.15 envelope). **Deviation from R for multi-character baseline states (D >= 10 or negative D):** R's `did_multiplegt_by_path` derives the per-path baseline via `path_index$baseline_XX <- substr(path_index$path, 1, 1)`, which captures only the first character of the comma-separated path string. For multi-character baselines this drops the rest of the value: for `path = "12,12,..."` it captures `"1"` instead of `"12"`; for `path = "-1,-1,..."` it captures `"-"` instead of `"-1"`. R's per-path control-pool subset is mis-allocated in both regimes. Python's tuple-key matching is correct — the per-path point estimates we compute are correct; R's per-path subset for the same path is buggy. The shipped R-parity scenarios stay in nonnegative single-digit `D in {0, 1, 2}` to avoid the R bug; negative-integer baseline support is regression-tested in Python only (no R parity). R-parity test at `tests/test_chaisemartin_dhaultfoeuille_parity.py::TestDCDHDynRParityByPathNonBinary`; cross-surface invariants regression-tested at `tests/test_chaisemartin_dhaultfoeuille.py::TestByPathNonBinary`.
+- **`ChaisemartinDHaultfoeuille.by_path` + non-binary integer treatment** — `by_path=k` now accepts integer-coded discrete treatment (D in Z, e.g. ordinal `{0, 1, 2}`); path tuples become integer-state tuples like `(0, 2, 2, 2)`. The previous `NotImplementedError` gate at `chaisemartin_dhaultfoeuille.py:1870` is replaced by a `ValueError` for continuous D (e.g. `D=1.5`) at fit-time per the no-silent-failures contract — the existing `int(round(float(v)))` cast in `_enumerate_treatment_paths` is now defensive (no-op for integer-coded D). Validated against R `did_multiplegt_dyn(..., by_path)` for D in `{0, 1, 2}` via the new `multi_path_reversible_by_path_non_binary` golden-value scenario (78 switchers, 3 paths, single-baseline custom DGP, F_g >= 4): per-path point estimates match R bit-exactly (rtol ~1e-9 on event horizons; rtol+atol envelope for placebo near-zero values), per-path SE inherits the documented cross-path cohort-sharing deviation (~5% rtol observed; SE_RTOL=0.15 envelope). **Deviation from R for multi-character baseline states (D >= 10 or negative D):** R's `did_multiplegt_by_path` derives the per-path baseline via `path_index$baseline_XX <- substr(path_index$path, 1, 1)`, which captures only the first character of the comma-separated path string. For multi-character baselines this drops the rest of the value: for `path = "12,12,..."` it captures `"1"` instead of `"12"`; for `path = "-1,-1,..."` it captures `"-"` instead of `"-1"`. R's per-path control-pool subset is mis-allocated in both regimes. Python's tuple-key matching is correct — the per-path point estimates we compute are correct; R's per-path subset for the same path is buggy. The shipped R-parity scenarios stay in nonnegative single-digit `D in {0, 1, 2}` to avoid the R bug; negative-integer treatment-state support (paths containing negative D values in non-baseline positions) is regression-tested in Python only at `tests/test_chaisemartin_dhaultfoeuille.py::TestByPathNonBinary::test_negative_integer_D_supported` (no R parity); a dedicated regression for a negative-baseline path (e.g. `(-1, 0, 0, 0)`) is deferred. R-parity test at `tests/test_chaisemartin_dhaultfoeuille_parity.py::TestDCDHDynRParityByPathNonBinary`; cross-surface invariants regression-tested at `tests/test_chaisemartin_dhaultfoeuille.py::TestByPathNonBinary`.
 - **New `paths_of_interest` kwarg on `ChaisemartinDHaultfoeuille`** for user-specified treatment-path subsets, alternative to `by_path=k`'s top-k automatic ranking. Mutually exclusive with `by_path`; setting both raises `ValueError` at `__init__` and `set_params` time. Each path tuple must be a list/tuple of `int` of length `L_max + 1` (uniformity validated at `__init__`; length match against `L_max + 1` validated at fit-time); `bool` and `np.bool_` are explicitly rejected, `np.integer` accepted and canonicalized to Python `int` for tuple-key consistency. Duplicates emit a `UserWarning` and are deduplicated; paths not observed in the panel emit a `UserWarning` and are omitted from `path_effects`. Paths appear in `results.path_effects` in the user-specified order, modulo deduplication and unobserved-path filtering. Composes with non-binary D and all downstream `by_path` surfaces (bootstrap, per-path placebos, per-path joint sup-t bands, `controls`, `trends_linear`, `trends_nonparam`) — mechanical filter on observed paths via the same `_enumerate_treatment_paths` call site, no methodology change. **Python-only API extension; no R equivalent** — R's `did_multiplegt_dyn(..., by_path=k)` only accepts a positive int (top-k) or `-1` (all paths). The `by_path` precondition gate at `chaisemartin_dhaultfoeuille.py:1118` (drop_larger_lower / L_max / `heterogeneity` / `design2` / `honest_did` / `survey_design` mutex) and the 11 `self.by_path is not None` activation branches in `fit()` were rerouted to fire under either selector. Validation + behavior + cross-feature regressions at `tests/test_chaisemartin_dhaultfoeuille.py::TestPathsOfInterest`.
 - **HAD `practitioner_next_steps()` handler + `llms-full.txt` reference section** (Phase 5). Adds `_handle_had` and `_handle_had_event_study` to `diff_diff/practitioner.py::_HANDLERS`, routing both `HeterogeneousAdoptionDiDResults` (single-period) and `HeterogeneousAdoptionDiDEventStudyResults` (event-study) through HAD-specific Baker et al. (2025) step guidance: `did_had_pretest_workflow` (step 3 — paper Section 4.2 step-2 closure on the event-study path), an estimand-difference routing nudge to `ContinuousDiD` (step 4 — fires when the user wants per-dose ATT(d) / ACRT(d) curves rather than HAD's WAS estimand and has never-treated controls; framed around estimand difference, NOT around the existence of untreated units, since HAD remains valid with a small never-treated share per REGISTRY § HeterogeneousAdoptionDiD edge cases and explicitly retains never-treated units on the staggered event-study path per paper Appendix B.2 / `had.py:1325`), `results.bandwidth_diagnostics` inspection on continuous designs and simultaneous (sup-t) `cband_*` reading on weighted event-study fits (step 6), per-horizon WAS event-study disaggregation (step 7), and the explicit design-auto-detection / last-cohort-only-WAS framing (step 8). Symmetric pair: `_handle_continuous` gains a Step-4 nudge to `HeterogeneousAdoptionDiD` for ContinuousDiD users on no-untreated panels (this direction is correct because ContinuousDiD's identification requires never-treated controls). Extends `_check_nan_att` with an ndarray branch via lazy `numpy` import for HAD's per-horizon `att` array; uses `np.all(np.isnan(arr))` semantics so partial-NaN arrays (legitimate event-study output under degenerate horizon-specific designs) do not over-fire the warning. Scalar path is bit-exact preserved across all 12 untouched handlers. Adds full HAD section + `HeterogeneousAdoptionDiDResults` / `HeterogeneousAdoptionDiDEventStudyResults` blocks + `## HAD Pretests` index covering all 7 pretest entry points + Choosing-an-Estimator row to `diff_diff/guides/llms-full.txt` (the bundled-in-wheel agent reference); the documented constructor + `fit()` signatures match the real `HeterogeneousAdoptionDiD.__init__` / `.fit` API exactly (verified by `inspect.signature`-based regression tests). Tightens the existing `Continuous treatment intensity` Choosing row to surface ATT(d) vs WAS as the estimand differentiator. `docs/doc-deps.yaml` updated to remove the `llms-full.txt` deferral note on `had.py` and add `llms-full.txt` entries to `had.py`, `had_pretests.py`, and `practitioner.py` blocks. Patch-level (additive on stable surfaces). 26 new tests (16 in `tests/test_practitioner.py::TestHADDispatch` + 9 in `tests/test_guides.py::TestLLMsFullHADCoverage` + 1 fixture-minimality regression locking the "handlers are STRING-ONLY at runtime" stability invariant). Closes the Phase 5 "agent surfaces" gap. T21 pretest tutorial subsequently landed in PR #409; T22 weighted/survey tutorial remains queued as a separate notebook PR.
 

TODO.md

@@ -59,6 +59,7 @@ Deferred items from PR reviews that were not addressed before merge.
 | dCDH: Phase 1 per-period placebo DID_M^pl has NaN SE (no IF derivation for the per-period aggregation path). Multi-horizon placebos (L_max >= 1) have valid SE. | `chaisemartin_dhaultfoeuille.py` | #294 | Low |
 | dCDH: Survey cell-period allocator's post-period attribution is a library convention, not derived from the observation-level survey linearization. MC coverage is empirically close to nominal on the test DGP; a formal derivation (or a covariance-aware two-cell alternative) is deferred. Documented in REGISTRY.md survey IF expansion Note. | `chaisemartin_dhaultfoeuille.py`, `docs/methodology/REGISTRY.md` | PR 2 | Medium |
 | dCDH: Parity test SE/CI assertions only cover pure-direction scenarios; mixed-direction SE comparison is structurally apples-to-oranges (cell-count vs obs-count weighting). | `test_chaisemartin_dhaultfoeuille_parity.py` | #294 | Low |
+| dCDH by_path: negative-baseline path regression (e.g. `(-1, 0, 0, 0)`) is not yet exercised. The existing negative-D test (`test_negative_integer_D_supported`) only covers paths with negative values in non-baseline positions like `(0, -1, -1, -1)`, which does not trigger the R `substr(path, 1, 1)` bug regime (the bug needs a multi-character baseline). Add a switcher fixture with `D_{g,1} = -1` and assert the resulting path tuple key. | `tests/test_chaisemartin_dhaultfoeuille.py` | #419 | Low |
 | CallawaySantAnna: consider materializing NaN entries for non-estimable (g,t) cells in group_time_effects dict (currently omitted with consolidated warning); would require updating downstream consumers (event study, balance_e, aggregation) | `staggered.py` | #256 | Low |
 | ImputationDiD dense `(A0'A0).toarray()` scales O((U+T+K)^2), OOM risk on large panels | `imputation.py` | #141 | Medium (deferred — only triggers when sparse solver fails) |
 | Multi-absorb weighted demeaning needs iterative alternating projections for N > 1 absorbed FE with survey weights; unweighted multi-absorb also uses single-pass (pre-existing, exact only for balanced panels) | `estimators.py` | #218 | Medium |