Add LLM-based decision point evaluator skill and supporting library tooling

[ahouseholder]

Summary

Add an SSVC domain skill (evaluate-decision-point) and reusable library tooling that enables an LLM agent to evaluate a single SSVC decision point given caller-assembled evidence, producing a validated Selection object as structured output.

Motivation

SSVC decision point evaluation is not a free-form prose task — it produces structured semantic objects (Selection, SelectionList) that must conform to the existing Pydantic schema in src/ssvc/selection.py. An LLM-based evaluator needs a typed validation boundary between probabilistic inference and deterministic orchestration. Without this, evaluation results are unverified prose that cannot be consumed programmatically.

Core idea

The key abstraction is:

• Evidence in (caller-assembled: text, structured JSON, CVE records, advisories, links, etc.)
• Validated Selection out (conforming to ssvc.selection.Selection)

The evaluator should live in src/ssvc/ as a reusable library component. An SSVC domain skill (skills/ssvc/evaluate-decision-point/) would invoke the library component and report results. The skill and library should be designed so the caller is responsible for assembling and formatting input evidence; the library provides structure and output validation, not input parsing.

The initial implementation is expected to use pydantic-ai as the agent framework (it provides schema validation, typed outputs, retry semantics, and parsing enforcement for structured LLM outputs), but the design should separate design intent from implementation choice.

Scope

• One decision point evaluated per invocation
• Caller-assembled evidence (no mandatory input format)
• Library provides flexible structure for callers to build on
• Skill orchestrates the library call and reports back a Selection
• Failure modes (invalid schema emission, hallucinated fields, unsupported values, malformed selections) should be recognized and handled via retry semantics — but the exact mechanisms are deferred to implementation

Out of scope (for now)

• Multi-decision-point batch evaluation (full SelectionList in one pass)
• Schema extensions to Selection/SelectionList for rationale, uncertainty, or provenance (TBD — may need a wrapper or extension; unclear)
• Model provider selection/configuration
• Input parsing (caller is responsible for assembling evidence)

Dependencies

Blocked by #1151 (skills infrastructure: two-tier directory structure and canonical SKILL.md format must exist before this skill can be added properly).

Notes

An earlier exploratory sketch (docs/exploratory/llm_evaluator.md on feature/pydantic.ai) describes the high-level flow; that branch is defunct but the flowchart captures intent. The existing Selection / SelectionList Pydantic models (src/ssvc/selection.py) are already the natural output type for this evaluator.