feat: Add infer_schema with confidence-scored type inference (#855)

[dynamo-pentester]

Summary

Implements automatic schema/data type inference with confidence scoring, as scoped in #855.

Adds infer_schema(frame: ArFrame) -> InferredSchema to the quality layer. For each column, it returns a ColumnInference with a best-guess inferred_type, a deterministic confidence score in [0.0, 1.0], an is_ambiguous flag, and the full candidates score breakdown.

Closes #855

What changed

• arnio/quality.py

  • Added ColumnInference and InferredSchema frozen dataclasses
  • Added infer_schema(frame) — a thin layer over the existing profile() output (reuses suggested_dtype, semantic_type, null_ratio, unique_ratio; no second parsing engine)
  • Defined _INFER_CANDIDATE_TYPES (the six supported types: int64, float64, bool, datetime, categorical, string) and _AMBIGUITY_THRESHOLD = 0.15 in one place
  • ColumnInference.to_dict() / InferredSchema.to_dict() — fully JSON-safe, deterministic column/candidate ordering
  • InferredSchema.to_schema() — maps only to the dtypes Field/Schema already accept (int64, float64, bool, datetime, string); categorical → string. All fields default to nullable=True (conservative default)
  • Added from .schema import Field, Schema import for the to_schema() return type

• arnio/__init__.py

  • Exported ColumnInference, InferredSchema, infer_schema

• tests/test_infer_schema.py (new)

  • Primitive type detection (int64, float64, bool, datetime, string)
  • Boolean variants: yes/no, true/false, 1/0
  • Multiple datetime formats
  • Mixed numeric/string columns
  • High-null and all-null columns (predictable fallback, no misleading high confidence)
  • Ambiguity boundary (second candidate within 0.15 of top)
  • Categorical detection (documented rule: ≥2 distinct values AND unique_ratio ≤ 0.20 — not uniqueness alone)
  • Score validation: all confidence/candidate scores finite and in [0.0, 1.0]
  • Deterministic output across repeated calls
  • to_dict() JSON serialization and ordering
  • to_schema() → valid Schema, correct dtype mappings, usable with ar.validate()
  • Input validation (TypeError on non-ArFrame)

• website/api.html

  • Added infer_schema(frame) entry to the Quality section, documenting ColumnInference fields and InferredSchema methods

Out of scope (untouched)

• profile(), compare_profiles(), detect_drift() / DriftReport
• Schema, Field, validators (only consumed via to_schema())
• C++ core, CSV parser, pipeline

How to test

pip install -e ".[dev]"
pytest tests/test_infer_schema.py -v
pytest tests/ -q   # full suite, ensure no regressions

Example usage

import arnio as ar

frame = ar.read_csv("data.csv")
schema = ar.infer_schema(frame)

for name, col in schema.columns.items():
    print(name, col.inferred_type, f"{col.confidence:.2f}", "ambiguous" if col.is_ambiguous else "")

# Pipe directly into validate()
result = ar.validate(frame, schema.to_schema())

[vercel]

@dynamo-pentester is attempting to deploy a commit to the xtylishanish-gmailcom's projects Team on Vercel.

A member of the Team first needs to authorize it.