…d QRF

model.py: the ConditionalModel/FittedModel protocols and resolve_fit_weights — the single authority enforcing that a fit is weighted by construction (weights reads the owning entity's typed Weights; weights='none' is the only unweighted path; a misspelled or mismatched kind raises rather than silently fitting unweighted). predictors_targets_entity refuses predictors/targets that span entities.

qrf.py: the canonical model. Regime detection (structural/unweighted sign-support gates), sequential chaining (each target conditions on predictors plus the targets already drawn), and weighted bootstrap (importance-resample rows by weight before growing each forest — the microimpute#196 fix, reimplemented from scratch against the Frame, not imported). Draws sample the weighted conditional by querying the forest at a per-row seeded quantile.

__init__.py: public API (ConditionalModel, QRF/RegimeGatedQRF, fit) and the constellation compatibility gate — asserts populace-frame's major/minor at import so a loose resolver cannot assemble an incompatible kernel pair.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…d placeholder)

test_weighted_fit_contract.py is the real realization of the kernel's skipped test_weighted_fit_shifts_draws_toward_weighted_truth: on a donor whose target is large exactly where weight is small (the #196 shape, ~20% low-weight huge-value rows), the weighted fit's draws land within 20% of the true weighted mean, weights='none' lands within 20% of the unweighted mean, and the two differ by >3x. Also asserts the default is weighted (no unweighted default).

test_qrf.py: regime gates preserve a zero-inflated target's zero mass and both signs with no zero-crossing; chaining reproduces a cross-target correlation; predict row-count/index match the input; fixed seed is deterministic; successive predicts draw independently.

test_model.py: weights='none' is the only unweighted path (a typo'd kind raises and names it; a mismatched kind raises and names the stored kind); predictors/targets spanning entities are refused. test_compat.py exercises the import gate.

n=5000 seeded for CI speed; 35 tests.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

resolve_fit_weights now resolves via frame.resolve_weights(entity) rather
than frame.weights_for(entity), so a person-level fit on a
household-weighted frame inherits the household weights (and their kind)
through membership instead of raising. This was the bug: the canonical CPS
shape (person predictors/targets, design weights only on the household)
could not be fit weighted at all.

The kind discipline is unchanged — the requested kind must match the
resolved (possibly inherited) kind, else raise. The impossible-remediation
message is fixed: requesting "design" on a calibrated frame no longer
advises "advance the frame's weights to design" (kinds only move forward,
so that is impossible); it now tells the caller to pass
weights="calibrated", the kind the frame actually carries. The forward
direction (e.g. requesting calibrated on a design frame) keeps the
advance-the-weights advice.

Regression tests (test_model.py): the CPS shape fits weighted and the
resolved vector broadcasts the household weights onto persons; a default
design fit on a calibrated frame raises naming weights="calibrated" and
not the impossible advance-to-design advice. The existing kind-mismatch
test now matches "resolved weights" wording.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The sign-class gate (HistGradientBoostingClassifier) was fit on an n-of-n
weighted bootstrap, which deletes vanishingly-rare low-weight classes
outright: a positive row at weight 1 among thousands of zeros at weight 50
is drawn with probability ~4e-5, so the resampled labels routinely contain
only the zero class and the gate could never draw the positive sign (0
positive draws across millions, reproduced).

HistGradientBoostingClassifier honors sample_weight exactly, so the gate
is now fit with sample_weight=weights directly, no bootstrap. Every
training row is present, so every sign class the data contains survives
into classes_. The weighted bootstrap stays for the QRF forests, which
genuinely need it: quantile-forest uses sample_weight only as a >0 leaf
mask (confirmed _quantile_forest.py:266), so it ignores weight magnitude
and the resample is the only way to weight the leaf distributions.

A guard now enforces internal consistency: if a sign class present in the
training labels is absent from the fitted gate's classes_, the fit raises
rather than silently drawing that class at probability zero.

Regression tests (test_qrf.py): the reviewer's repro (n=5000, ~10 positive
at weight 1, ~4990 zero at weight 50) keeps both gate classes and produces
positive draws across seeds (was 0/2M); the consistency guard raises when
a stubbed gate drops a training class.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…ndpoints, chunking)

Four changes to how the quantile forest is grown and read out, so draws
reproduce the conditional's tail instead of undershooting it ~2x:

1. max_samples_leaf=None. The forests are now grown keeping ALL leaf
   samples; quantile-forest's default of 1 keeps one sample per leaf,
   thinning each row's conditional to ~n_estimators atoms and undershooting
   tail mass. Exposed as a RegimeGatedQRF param (default None). On the
   contract fixture the weighted share above 300k goes from ~0.0035
   (nearest-snap, msl=1) to ~0.0050 — the weighted-population truth.

2. Linear interpolation. draw() no longer snaps each row to the nearest of
   201 grid points (which quantizes every draw and biases the tail toward
   the bracket interior); it linearly interpolates the row's value at its
   exact quantile between the two bracketing grid quantiles.

3. Drawable extremes. The quantile grid now includes points adjacent to 0
   and 1, so the observed conditional min and max are drawable. q=1 is the
   observed maximum, not extrapolation — the old comment wrongly excluded
   it. With a lone extreme the interior-only grid (top q=0.995) reads far
   below the max; the endpoint reaches it.

4. Chunked predict. draw() batches the predict over rows
   (_PREDICT_CHUNK_ROWS=50k) so the (n_rows x n_grid) matrix never
   materializes whole — at 3M+ rows it would be tens of GB. Chunking is
   bit-identical to a single pass (quantiles are drawn up front and sliced
   positionally).

Regression tests (test_qrf.py): the weighted tail share above 300k is
within ~2x of truth and materially closer than the nearest-snap baseline;
a draw at q->1 reaches the observed conditional max via the grid endpoint,
which the interior-only winsorized grid misses.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

NaN targets were silently relabeled to the zero class: the sign labels
(y > atol / y < -atol) are both False for NaN, so a missing value was
miscoded as a structural zero, NaN-blind. The model has no notion of
missingness, so fit now validates at entry that every target column is
entirely finite and raises a ValueError naming the offending column and
its non-finite count (NaN or inf). Predictors are not checked — a forest
splits around NaN features and a missing predictor is not silently
miscoded the way a missing target is.

Regression tests (test_qrf.py): a target with 3 NaNs raises naming the
column and the count; an inf target is refused the same way.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…tract

Finding F: the fitted model was seeded with the raw model seed, making its
draw uniforms bit-identical to the fit's bootstrap-selection uniforms (the
draws were not independent of the fit's resampling). Seed fit and draw from
two independent SeedSequence children of the model seed; determinism is
preserved. Finding H: add a contract that the zero gate reproduces the
*weighted* (population) zero-share, not the sample's, when the two differ.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…olumns

Finding G: pin populace-frame>=0.1,<0.2 (the constellation must resolve, not
fail only at the import-time compat gate), and refuse duplicate predictors,
duplicate targets, or a column that is both predictor and target (these
silently fit twice / fit P(y|y) before). Messages name the culprits.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…ification review)

Two independent verification reviews (mutation testing) found that the
resolve_weights kind fix introduced two bugs in _inherited_kind, which
only handled the weighted-group path while _effective_weights gives
person-stored weights precedence when deriving a group entity:

1. Regression: weighted accounting (wsum/wmean/wquantile/gini/...) of a
   group-entity column on a person-only-weighted frame raised instead of
   deriving the group weights from the person weights — on a frame shape
   the fit suite's own fixtures build.
2. Silent kind mislabel: a third entity's resolved values (from the
   person source) were tagged with a sibling group's kind, which could
   leak through resolve_fit_weights as a kind-discipline violation.

Make _inherited_kind recurse exactly as _effective_weights does so kind
always names the source the values come from. Three regression tests:
person-only group accounting, mixed-kind coherence, and a leaf-component
pin for the tail-draw fix (max_samples_leaf=None was untested).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>