feat(runner): classify privacy through llm targets

[afourniernv]

Draft stack 7/10.

What

Allow the semantic privacy classifier to use either a typed decision target or an ordinary LLM target.

The LLM form receives the same bounded request context and must return the configured structured verdict.

Why

Typed decisions are the tighter integration, but some deployments need to use a model already available through Switchyard's LLM clients.

How tested

Tests cover both classifier types, valid and invalid verdicts, failures, target isolation, required timeouts, and unsupported request modifiers. The complete stack passed workspace formatting, Clippy with warnings denied, and workspace tests. No live provider calls.

Notes for reviewers

The typed decision classifier remains the default. An LLM classifier is selected with:

[routes.<name>.privacy.classifier]
type = "llm"
target = "privacy_classifier"
clear_threshold = 0.9

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.