views-frames 1.8.1__tar.gz → 1.10.0__tar.gz

{views_frames-1.8.1 → views_frames-1.10.0}/.gitignore

@@ -44,3 +44,6 @@ research/**/.cache/
 research/**/results.tsv
 research/**/run.log
 research/**/audit_plots/
+
+# generated figures (research/figures/ generators; deliverable copies travel with the deliverable)
+reports/plots/

{views_frames-1.8.1 → views_frames-1.10.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: views-frames
-Version: 1.8.1
+Version: 1.10.0
 Summary: The VIEWS platform data-contract layer: immutable array+identifier frames (numpy only, root of the dependency DAG).
 Project-URL: Homepage, https://github.com/views-platform/views-frames
 Project-URL: Repository, https://github.com/views-platform/views-frames
@@ -36,7 +36,7 @@ Description-Content-Type: text/markdown
 > containers (`FeatureFrame`, `PredictionFrame`, and their anticipated siblings)
 > that every other repo depends on and that depends on nothing internal.
 >
-> **Status:** **v1.8.0 — frozen API, published to PyPI** (frozen since v1.0.0, ADR-018; the
+> **Status:** **v1.10.0 — frozen API, published to PyPI** (frozen since v1.0.0, ADR-018; the
 > v1.1 surface is
 > purely additive — the coherent posterior summary, ADR-019; v1.2.0 rebuilt the tower
 > `outside-in`, C-44; v1.3.0 makes the tower summary distribution-agnostic — no magnitude
@@ -47,7 +47,13 @@ Description-Content-Type: text/markdown
 > sibling package **`views_frames_reconcile`** — forecast reconciliation, ADR-023; v1.8.0
 > adds the native point-country broadcast + the self-describing reconciliation mode,
 > the three showcase notebooks, and the frames-family hardening pass — Reconcile.md CIC,
-> ADR-025 immutability-by-convention, the adversarial red-test batch). This
+> ADR-025 immutability-by-convention, the adversarial red-test batch; v1.9.0 moves the
+> tower-tip MAP to the **top floor** of the published tower — `tip_mass` 0.25 with the
+> **MAP-containment law** in the conformance suite, ADR-019 Amendment 3 — and adds the
+> `research/figures/` tower-figure toolkit; v1.10.0 adds the **dense-grid fill**
+> primitive — `reindex_fill(other, *, fill_value)` on all three frames +
+> `SpatioTemporalIndex.cartesian` + the published `assert_reindex_fill_law`, ADR-026,
+> unblocking pandas-free FAO ingestion). This
 > README is the design
 > bible; the contract it specifies is realised in `src/views_frames/` (index, frames,
 > io, conformance suite) plus the `src/views_frames_summarize/` sibling package
@@ -343,7 +349,12 @@ genuinely reused core. Build this once:
   `(time, unit)` **at a single `SpatialLevel`**. **This is the label-alignment
   that today drags pandas back in** — pred↔actual join, partial-overlap
   evaluation, same-level reindex. This alignment logic lives in the leaf
-  unconditionally.
+  unconditionally. The dense-grid pair rides the same join (ADR-026):
+  `cartesian(times, units, level)` builds the time-major product index
+  (explicit arrays only — deriving them is consumer policy), and the frames'
+  `reindex_fill(other, *, fill_value)` aligns with **no** superset requirement,
+  filling absent rows (present rows bit-exact; law-pinned by
+  `assert_reindex_fill_law`).
 - **Cross-level operations (`cross_level_align`) — protocol here, data injected.**
   The cm↔pgm **cross-level join** (country↔grid) is **not** a same-axis set op; it
   is a one-to-many lookup against a `priogrid→country` mapping that is **injected**

{views_frames-1.8.1 → views_frames-1.10.0}/README.md

@@ -4,7 +4,7 @@
 > containers (`FeatureFrame`, `PredictionFrame`, and their anticipated siblings)
 > that every other repo depends on and that depends on nothing internal.
 >
-> **Status:** **v1.8.0 — frozen API, published to PyPI** (frozen since v1.0.0, ADR-018; the
+> **Status:** **v1.10.0 — frozen API, published to PyPI** (frozen since v1.0.0, ADR-018; the
 > v1.1 surface is
 > purely additive — the coherent posterior summary, ADR-019; v1.2.0 rebuilt the tower
 > `outside-in`, C-44; v1.3.0 makes the tower summary distribution-agnostic — no magnitude
@@ -15,7 +15,13 @@
 > sibling package **`views_frames_reconcile`** — forecast reconciliation, ADR-023; v1.8.0
 > adds the native point-country broadcast + the self-describing reconciliation mode,
 > the three showcase notebooks, and the frames-family hardening pass — Reconcile.md CIC,
-> ADR-025 immutability-by-convention, the adversarial red-test batch). This
+> ADR-025 immutability-by-convention, the adversarial red-test batch; v1.9.0 moves the
+> tower-tip MAP to the **top floor** of the published tower — `tip_mass` 0.25 with the
+> **MAP-containment law** in the conformance suite, ADR-019 Amendment 3 — and adds the
+> `research/figures/` tower-figure toolkit; v1.10.0 adds the **dense-grid fill**
+> primitive — `reindex_fill(other, *, fill_value)` on all three frames +
+> `SpatioTemporalIndex.cartesian` + the published `assert_reindex_fill_law`, ADR-026,
+> unblocking pandas-free FAO ingestion). This
 > README is the design
 > bible; the contract it specifies is realised in `src/views_frames/` (index, frames,
 > io, conformance suite) plus the `src/views_frames_summarize/` sibling package
@@ -311,7 +317,12 @@ genuinely reused core. Build this once:
   `(time, unit)` **at a single `SpatialLevel`**. **This is the label-alignment
   that today drags pandas back in** — pred↔actual join, partial-overlap
   evaluation, same-level reindex. This alignment logic lives in the leaf
-  unconditionally.
+  unconditionally. The dense-grid pair rides the same join (ADR-026):
+  `cartesian(times, units, level)` builds the time-major product index
+  (explicit arrays only — deriving them is consumer policy), and the frames'
+  `reindex_fill(other, *, fill_value)` aligns with **no** superset requirement,
+  filling absent rows (present rows bit-exact; law-pinned by
+  `assert_reindex_fill_law`).
 - **Cross-level operations (`cross_level_align`) — protocol here, data injected.**
   The cm↔pgm **cross-level join** (country↔grid) is **not** a same-axis set op; it
   is a one-to-many lookup against a `priogrid→country` mapping that is **injected**

{views_frames-1.8.1 → views_frames-1.10.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "views-frames"
-version = "1.8.1"
+version = "1.10.0"
 description = "The VIEWS platform data-contract layer: immutable array+identifier frames (numpy only, root of the dependency DAG)."
 authors = [
     { name = "Simon Polichinel von der Maase", email = "simmaa@prio.org" },

{views_frames-1.8.1 → views_frames-1.10.0}/src/views_frames/conformance/__init__.py

@@ -37,6 +37,7 @@ __all__ = [
     "assert_frame_contract",
     "assert_frame_envelope",
     "assert_index_alignment_laws",
+    "assert_reindex_fill_law",
 ]
 
 
@@ -136,6 +137,50 @@ def assert_index_alignment_laws(index_a: Any, index_b: Any) -> None:
     assert np.array_equal(index_a.unit[pos], index_a.unit), "searchsorted self-identity"
 
 
+def assert_reindex_fill_law(frame: Any, target: Any, fill_value: float) -> None:
+    """Assert the dense-grid fill law (ADR-026) for ``frame`` against ``target``.
+
+    ``reindex_fill`` aligns a frame to a target index with **no** superset
+    requirement, filling absent rows. The law:
+
+    - the result's index equals ``target`` row-for-row (time, unit, level);
+    - every target row present in ``frame`` comes through **bit-exact**;
+    - every absent row equals ``fill_value`` exactly (NaN-safe comparison);
+    - when every target row is present, ``reindex_fill`` degenerates to
+      ``reindex`` (filling a superset frame adds nothing).
+
+    Args:
+        frame: a frame exposing ``reindex_fill``/``reindex``/``index``/``values``.
+        target: a ``SpatioTemporalIndex`` (same level) to densify against.
+        fill_value: the fill for absent rows (``NaN`` is legal).
+
+    Raises:
+        AssertionError: a law is violated.
+    """
+    _require_assertions()
+    filled = frame.reindex_fill(target, fill_value=fill_value)
+    assert filled.index.level == target.level, "reindex_fill must keep the level"
+    assert np.array_equal(filled.index.time, target.time), (
+        "reindex_fill result index must equal the target (time)"
+    )
+    assert np.array_equal(filled.index.unit, target.unit), (
+        "reindex_fill result index must equal the target (unit)"
+    )
+    pos = frame.index.searchsorted(target)
+    found = pos >= 0
+    assert np.array_equal(
+        filled.values[found], frame.values[pos[found]], equal_nan=True
+    ), "reindex_fill must pass present rows through bit-exact"
+    expected = np.full_like(filled.values[~found], np.float32(fill_value))
+    assert np.array_equal(filled.values[~found], expected, equal_nan=True), (
+        "reindex_fill must set every absent row to fill_value"
+    )
+    if bool(found.all()):
+        assert np.array_equal(
+            filled.values, frame.reindex(target).values, equal_nan=True
+        ), "reindex_fill on a superset frame must equal reindex"
+
+
 def assert_cross_level_alignment_law(
     index: Any, mapping: Any, target_level: Any
 ) -> None:

{views_frames-1.8.1 → views_frames-1.10.0}/src/views_frames/feature_frame.py

@@ -154,6 +154,37 @@ class FeatureFrame:
             )
         return self.select(self._index.searchsorted(other))
 
+    def reindex_fill(
+        self, other: SpatioTemporalIndex, *, fill_value: float
+    ) -> FeatureFrame:
+        """Align to ``other``'s rows, filling rows absent from this frame.
+
+        The dense-grid companion to :meth:`reindex` (ADR-026): every row of
+        ``other`` present in this frame comes through **bit-exact**; a row
+        absent from it gets ``fill_value``, broadcast across the feature and
+        sample axes. There is no superset requirement — that is the difference
+        from the fail-loud ``reindex``. The result's index **is** ``other``
+        (immutable, shared); ``feature_names`` and metadata are preserved.
+
+        Same-level join semantics: assumes unique ``(time, unit)`` rows in
+        **this** frame (register C-21); ``other`` may repeat rows. Allocates
+        the full ``(other.n_rows, F, S)`` buffer — densifying a large grid is
+        a deliberate, costly act. A ``NaN`` fill is legal here, but the
+        summarize estimators fail loud on NaN draws by design (ADR-017).
+
+        Raises:
+            ValueError: ``other`` is at a different ``SpatialLevel``.
+        """
+        pos = self._index.searchsorted(other)
+        found = pos >= 0
+        out = np.full(
+            (other.n_rows, *self._values.shape[1:]),
+            np.float32(fill_value),
+            dtype=np.float32,
+        )
+        out[found] = self._values[pos[found]]
+        return FeatureFrame(out, other, self._feature_names, self._metadata)
+
     # ---- persistence --------------------------------------------------------
 
     def save(self, directory: Path | str) -> None:

{views_frames-1.8.1 → views_frames-1.10.0}/src/views_frames/index.py

@@ -43,9 +43,7 @@ class SpatioTemporalIndex:
         level: SpatialLevel,
     ) -> None:
         if not isinstance(level, SpatialLevel):
-            raise TypeError(
-                f"level must be a SpatialLevel, got {type(level).__name__}"
-            )
+            raise TypeError(f"level must be a SpatialLevel, got {type(level).__name__}")
         is_array = isinstance(time, np.ndarray) and time.ndim >= 1
         n = int(time.shape[0]) if is_array else -1
         validate_identifiers({"time": time, "unit": unit}, n_rows=n)
@@ -56,6 +54,65 @@ class SpatioTemporalIndex:
         self._unit.setflags(write=False)
         self._level = level
 
+    # ---- construction -------------------------------------------------------
+
+    @classmethod
+    def cartesian(
+        cls,
+        times: IntArray,
+        units: IntArray,
+        level: SpatialLevel,
+    ) -> SpatioTemporalIndex:
+        """The dense ``(time × unit)`` product index, in time-major order.
+
+        The dense-grid constructor (ADR-026): every combination of the given
+        ``times`` and ``units``, ordered time-major (all units of the first time,
+        then all units of the next), so dense indices are canonical across
+        consumers. The inputs are **explicit arrays only** — any rule for
+        deriving them (e.g. "the units present in the last time step") is
+        consumer policy and never lives in the leaf.
+
+        Fails loud on duplicated ``times`` or ``units``: a duplicated product
+        input manufactures duplicate ``(time, unit)`` rows, which make every
+        same-level join undefined (register C-21) — always a caller bug here.
+
+        Allocates ``len(times) × len(units)`` rows eagerly; densifying a large
+        grid is a deliberate, costly act (a full-pgm dense grid runs to tens of
+        millions of rows).
+
+        Args:
+            times: 1-D integer array of unique time identifiers.
+            units: 1-D integer array of unique unit identifiers.
+            level: The ``SpatialLevel`` of the produced index.
+
+        Raises:
+            TypeError: ``times``/``units`` is not an integer-dtype numpy array,
+                or ``level`` is not a ``SpatialLevel``.
+            ValueError: an input array is not 1-D or contains duplicate values.
+        """
+        for name, arr in (("times", times), ("units", units)):
+            if not isinstance(arr, np.ndarray):
+                raise TypeError(
+                    f"cartesian {name} must be a numpy array, got {type(arr).__name__}"
+                )
+            if not np.issubdtype(arr.dtype, np.integer):
+                raise TypeError(
+                    f"cartesian {name} must be an integer dtype, got {arr.dtype}"
+                )
+            if arr.ndim != 1:
+                raise ValueError(f"cartesian {name} must be 1-D, got ndim={arr.ndim}")
+            if np.unique(arr).shape[0] != arr.shape[0]:
+                raise ValueError(
+                    f"cartesian {name} contains duplicate values; a duplicated "
+                    "product input manufactures duplicate (time, unit) rows, "
+                    "which make same-level joins undefined (register C-21)"
+                )
+        return cls(
+            time=np.repeat(times, units.shape[0]),
+            unit=np.tile(units, times.shape[0]),
+            level=level,
+        )
+
     # ---- core surface -------------------------------------------------------
 
     @property
@@ -109,9 +166,11 @@ class SpatioTemporalIndex:
     @staticmethod
     def _row_view(keys: NDArray[np.int64]) -> NDArray[np.void]:
         """View each ``(time, unit)`` row as a single void scalar for set ops."""
-        return np.ascontiguousarray(keys).view(
-            np.dtype((np.void, keys.dtype.itemsize * keys.shape[1]))
-        ).reshape(-1)
+        return (
+            np.ascontiguousarray(keys)
+            .view(np.dtype((np.void, keys.dtype.itemsize * keys.shape[1])))
+            .reshape(-1)
+        )
 
     def _require_same_level(self, other: SpatioTemporalIndex) -> None:
         if self._level != other._level:
@@ -180,9 +239,7 @@ class SpatioTemporalIndex:
         rows = self._row_view(self._keys())
         return bool(len(np.unique(rows)) == rows.shape[0])
 
-    def select(
-        self, indexer: IntArray | NDArray[np.bool_]
-    ) -> SpatioTemporalIndex:
+    def select(self, indexer: IntArray | NDArray[np.bool_]) -> SpatioTemporalIndex:
         """A new index of the rows at integer positions **or** a boolean mask.
 
         The row-selection primitive the frame-level ``select``/``reindex`` build on:
@@ -298,9 +355,7 @@ class SpatioTemporalIndex:
         sorted_rows = map_rows[order]
 
         self_rows = self._row_view(self._keys())
-        pos = np.clip(
-            np.searchsorted(sorted_rows, self_rows), 0, len(sorted_rows) - 1
-        )
+        pos = np.clip(np.searchsorted(sorted_rows, self_rows), 0, len(sorted_rows) - 1)
         found = sorted_rows[pos] == self_rows
         if not bool(found.all()):
             miss = int(np.argmax(~found))

{views_frames-1.8.1 → views_frames-1.10.0}/src/views_frames/prediction_frame.py

@@ -118,6 +118,37 @@ class PredictionFrame:
             )
         return self.select(self._index.searchsorted(other))
 
+    def reindex_fill(
+        self, other: SpatioTemporalIndex, *, fill_value: float
+    ) -> PredictionFrame:
+        """Align to ``other``'s rows, filling rows absent from this frame.
+
+        The dense-grid companion to :meth:`reindex` (ADR-026): every row of
+        ``other`` present in this frame comes through **bit-exact**; a row
+        absent from it gets ``fill_value``, broadcast across the sample axis.
+        There is no superset requirement — that is the difference from the
+        fail-loud ``reindex``. The result's index **is** ``other`` (immutable,
+        shared); metadata is preserved.
+
+        Same-level join semantics: assumes unique ``(time, unit)`` rows in
+        **this** frame (register C-21); ``other`` may repeat rows. Allocates
+        the full ``(other.n_rows, S)`` buffer — densifying a large grid is a
+        deliberate, costly act. A ``NaN`` fill is legal here, but the
+        summarize estimators fail loud on NaN draws by design (ADR-017).
+
+        Raises:
+            ValueError: ``other`` is at a different ``SpatialLevel``.
+        """
+        pos = self._index.searchsorted(other)
+        found = pos >= 0
+        out = np.full(
+            (other.n_rows, *self._values.shape[1:]),
+            np.float32(fill_value),
+            dtype=np.float32,
+        )
+        out[found] = self._values[pos[found]]
+        return PredictionFrame(out, other, self._metadata)
+
     # ---- persistence --------------------------------------------------------
 
     def save(self, directory: Path | str) -> None:

{views_frames-1.8.1 → views_frames-1.10.0}/src/views_frames/target_frame.py

@@ -113,6 +113,36 @@ class TargetFrame:
             )
         return self.select(self._index.searchsorted(other))
 
+    def reindex_fill(
+        self, other: SpatioTemporalIndex, *, fill_value: float
+    ) -> TargetFrame:
+        """Align to ``other``'s rows, filling rows absent from this frame.
+
+        The dense-grid companion to :meth:`reindex` (ADR-026): every row of
+        ``other`` present in this frame comes through **bit-exact**; a row
+        absent from it gets ``fill_value``. There is no superset requirement —
+        that is the difference from the fail-loud ``reindex``. The result's
+        index **is** ``other`` (immutable, shared); metadata is preserved.
+
+        Same-level join semantics: assumes unique ``(time, unit)`` rows in
+        **this** frame (register C-21); ``other`` may repeat rows. Allocates
+        the full ``(other.n_rows, 1)`` buffer — densifying a large grid is a
+        deliberate, costly act. A ``NaN`` fill is legal here, but the
+        summarize estimators fail loud on NaN draws by design (ADR-017).
+
+        Raises:
+            ValueError: ``other`` is at a different ``SpatialLevel``.
+        """
+        pos = self._index.searchsorted(other)
+        found = pos >= 0
+        out = np.full(
+            (other.n_rows, *self._values.shape[1:]),
+            np.float32(fill_value),
+            dtype=np.float32,
+        )
+        out[found] = self._values[pos[found]]
+        return TargetFrame(out, other, self._metadata)
+
     # ---- persistence --------------------------------------------------------
 
     def save(self, directory: Path | str) -> None:

{views_frames-1.8.1 → views_frames-1.10.0}/src/views_frames_summarize/config.py

@@ -28,9 +28,15 @@ TOWER_CONFIG: Final[dict[str, Any]] = {
         [round(0.05 * i, 2) for i in range(1, 19)]  # 0.05 … 0.90
         + [0.92, 0.94, 0.95, 0.96, 0.97, 0.98, 0.99]  # fine high-mass tail
     ),
-    # The tower-tip reads the median of the floor at this mass — the "shorth". 0.5 is
-    # the maximally-robust choice: a duplicate would need ~half the draws to hijack it.
-    "tip_mass": 0.5,
+    # The tower-tip reads the median of the floor at this mass — the top floor of the
+    # published tower. 0.25 (the top-quartile floor; ADR-019 amendment 2026-07-24,
+    # evidence in research/map_hdi/tip_mass_study.py): lower bias than the 0.5 shorth
+    # everywhere, RMSE-superior at pooled S, exact on zero-inflated cells, and it
+    # passes the real-cell C-44 gate with margin — the tip floor's median is safe as
+    # long as a duplicate stack is under half the floor (k > 2·duplicates; at S=32,
+    # k=8 tolerates the observed 2-3-zero stacks). Containment law: every HDI of mass
+    # > tip_mass/2 provably contains the tip (asserted in conformance.py).
+    "tip_mass": 0.25,
     # Optional magnitude zero rule, OFF by default (``None``). When set to a float, a
     # row whose every draw is <= it collapses to 0 — a *count* opt-in (sub-1 ⇒ 0).
     # Left ``None`` for any non-count target: zero-inflation is otherwise handled by the

{views_frames-1.8.1 → views_frames-1.10.0}/src/views_frames_summarize/conformance.py

@@ -126,6 +126,32 @@ def _assert_tower_contract(frame: AnyFrame, n: int) -> None:
     assert (tip.values[..., 0] >= tlo - 1e-6).all(), "tip below the tip_mass floor"
     assert (tip.values[..., 0] <= thi + 1e-6).all(), "tip above the tip_mass floor"
 
+    # MAP-containment law (ADR-019 amendment 2026-07-24). Wider-than-tip_mass floors
+    # contain the tip by nesting. Below tip_mass, a nested floor still contains it
+    # whenever it holds MORE THAN HALF the tip floor's draws: a contiguous sub-window
+    # longer than half the parent cannot trim away the parent's middle draw(s), and
+    # the tip is their median/average. A floor of mass m spans floor(m·S)+1 draws
+    # (the `_ks` value counts inter-draw steps), so the exact condition is
+    # 2·(floor(m·S)+1) > floor(tip_mass·S)+1 — asymptotically mass > tip_mass/2.
+    # Floors below it carry NO guarantee and are below platform sample resolution
+    # (see tower_point.py / research/map_hdi/tip_mass_study.py).
+    s_count = int(frame.values.shape[-1])
+    n_tip = int(np.floor(tip_mass * s_count)) + 1
+    guaranteed = tuple(
+        float(m)
+        for m in config.canonical_floors()
+        if 2 * (int(np.floor(float(m) * s_count)) + 1) > n_tip
+    )
+    law_tower = hdi_tower(frame, masses=guaranteed)
+    for j, m in enumerate(guaranteed):
+        glo, ghi = law_tower[..., j, 0], law_tower[..., j, 1]
+        assert (tip.values[..., 0] >= glo - 1e-6).all(), (
+            f"MAP-containment violated: tip below the {m:.2f} floor"
+        )
+        assert (tip.values[..., 0] <= ghi + 1e-6).all(), (
+            f"MAP-containment violated: tip above the {m:.2f} floor"
+        )
+
     # Reproducibility law: the 50% HDI is independent of the other requested masses.
     just_50 = hdi_tower(frame, masses=(0.5,))
     assert np.array_equal(just_50[..., 0, :], tower[..., 0, :]), (

{views_frames-1.8.1 → views_frames-1.10.0}/src/views_frames_summarize/tower_point.py

@@ -2,15 +2,21 @@
 
 `tower_point` is the "most likely single value" we report to a consumer: the median of
 the draws inside the **`tip_mass` floor** of the nested tower (config-driven, default
-0.5 — the "shorth"). Zero-inflation is handled by that floor's density (a zero-majority
-row reads 0); the optional, off-by-default magnitude cutoff applies too if set (C-45).
+0.25 — the top-quartile floor, i.e. the top floor of the published tower; ADR-019
+amendment 2026-07-24, evidence in `research/map_hdi/tip_mass_study.py`). Zero-inflation
+is handled by that floor's density (a zero-majority row reads 0); the optional,
+off-by-default magnitude cutoff applies too if set (C-45). Containment law: every HDI
+whose floor holds more than half the tip floor's draws provably contains the tip
+(asymptotically, mass > `tip_mass`/2 — a nested contiguous window longer than half
+the parent cannot trim away the parent's median); narrower floors carry no guarantee
+and are below platform sample resolution. Asserted in ``conformance.py``.
 
 It is a *new* estimator, deliberately distinct from the frozen `map_estimate` (ADR-018):
 `map_estimate` is a binned histogram mode with a zero-*mass*-fraction rule and a
 lowest-index tie-break that is directionally biased on right-skewed, zero-inflated,
 low-sample posteriors (register C-32). The tower tip is unbinned — it reads the median
 of a mass-aware floor — so it carries no histogram tie-break **and** is robust to
-minority duplicated draws (register C-44; a lonely outlier cannot define a 50%-mass
+minority duplicated draws (register C-44; a lonely outlier cannot define a 25%-mass
 floor). On a genuinely multi-peaked row the tip is, like any point, ambiguous; pair it
 with `bimodality` to detect that case rather than collapse it silently.
 """