feat(runner): classify privacy through typed decisions

[afourniernv]

Draft stack 6/10.

What

Add an optional semantic privacy classifier backed by an existing typed decision target.

The classifier receives a bounded representation of the normalized request and returns a privacy verdict with a clear-content probability. Anything below the configured threshold is restricted.

Why

Deterministic checks cannot identify every confidential or regulated request. A typed decision model covers cases that require context while keeping the result constrained.

How tested

Tests cover bounded input, each verdict, invalid responses, provider failures, threshold behavior, and deterministic checks running before the classifier. The complete stack passed workspace formatting, Clippy with warnings denied, and workspace tests. No live provider calls.

Notes for reviewers

This reuses the existing RoutedDecisionClient and decision-target configuration.

Adds the privacy.classifier TOML section. No public Rust, Python, HTTP, or ABI changes.

[afourniernv]

Closing this draft stack for now. I’m sharing the cumulative branch and design/evaluation artifacts for architecture feedback before reopening a smaller stack.