Add standalone create_probs_histogram() prediction distribution adapt…

[uriahf]

…er and vendor rtichoke_viz v0.22.0

• Vendor rtichoke_viz v0.22.0 release asset (SHA-256 64087fe0284ab2beb6e664dd419504a56000e217eaf403f33559896b67fff94d).
• Implement _prediction_distribution_v2_spec_from_performance_data in _viz_spec_v2.py to construct canonical v0.22.0 PredictionDistributionSpec objects.
• Implement standalone public create_probs_histogram() function in probs_distribution.py and export in rtichoke package namespace.
• Update RtichokeBrowserChart to support single-file self-contained HTML rendering for prediction distribution charts.
• Add deterministic unit tests and Playwright real-browser acceptance tests.

[github-actions]

PR Preview Action v1.8.1
🚀 View preview at https://uriahf.github.io/rtichoke_python/pr-preview/pr-417/
Built to branch gh-pages at 2026-09-13 14:17 UTC. Preview will be ready when the GitHub Pages deployment is complete.

[uriahf]

Thanks — the overall implementation, v0.22.0 adoption, public API, join keys,
self-contained rendering, and test coverage look good. CI is green.

Before merging, please make one architectural correction so this consumer can
be reused cleanly by the upcoming Summary Report integration.

1. Split the current canonical builder into two layers.

The lowest-level shared builder should accept already-prepared inputs, along
the lines of:

_prediction_distribution_v2_spec(
    distribution_data,
    performance_data,
    evaluation_metadata,
)

It must only map those producer-owned quantities into the canonical
PredictionDistributionSpec. It must not call statistical producers.

The public standalone path may then use a convenience wrapper that calls:

_prepare_probs_distribution_data(...)
prepare_performance_data(...)
_build_evaluation_metadata(...)

and passes their outputs to the shared prepared-data builder.

The current function is named
_prediction_distribution_v2_spec_from_performance_data(), but it accepts raw
probs/reals and calculates performance internally. This prevents Summary
Reports from reusing their already-prepared threshold performance data and
would cause unnecessary repeated producer calls.

Do not change statistical behavior or the public create_probs_histogram() API.

2. Correct the PPCR wording in the public docstring.

Replace:

“Positive Predicted Value / Positive Rate”

with the established package terminology:

“Predicted Positives Condition Rate”

Risk Percentile may remain as the user-facing explanatory label.

3. Preserve confusion-matrix counts as Python integers.

Serialize true_positives, true_negatives, false_positives, and false_negatives
with int(...) rather than through the generic float conversion helper.
Continuous metrics should remain floats, and undefined estimates should remain
None/null.

Add focused tests proving:

• the prepared-data builder does not call either statistical producer;
• the public wrapper and prepared-data builder produce the same canonical spec;
• TP/TN/FP/FN estimates are integers;
• undefined continuous metrics remain None/null.

Keep Summary Report integration out of this PR. Do not merge yet.

[uriahf]

Superseded by #418, which contains the same standalone Prediction Distribution implementation together with the final published rtichoke_viz v0.22.1 artifact.