[TypeSpec Authoring] Add Vally assessment benchmarks

## Goal

Add a Vally benchmark suite that measures TypeSpec assessment routing and report quality across all five dimensions.

## Scope

- Add trigger and neighboring-skill anti-trigger cases for the assessment skill.
- Add curated capability cases for semantic intent, REST breaking changes, downstream SDK breaking changes, Azure compliance, and documentation.
- Include positive, no-impact, ambiguous, and blocked-evidence cases.
- Grade source-link accuracy, finding classification, intent-to-impact links, incomplete-analysis status, and absence of unsupported findings.
- Use hermetic fixtures for the PR gate; keep any live benchmark in a separate serialized tier.
- Record per-case scores, duration, and model so benchmark changes are comparable over time.

## Acceptance criteria

- The benchmark follows the repository Vally eval-authoring conventions and runs from the existing skill-eval pipeline.
- Routing coverage includes at least three triggers and three meaningful anti-triggers with competing skills mounted where applicable.
- Every report dimension has representative regression cases and explicit graders.
- Known no-impact cases penalize false positives, and blocked cases cannot pass as safe.
- Results identify the failed scenario and grader and can be compared with a checked-in baseline.
- Focused Vally evaluation passes locally and in CI.

Start by reading the repository's Vally eval-authoring conventions and the existing skill-eval pipeline. Define the hermetic fixtures, routing cases, capability cases, graders, and checked-in baseline around those conventions; done means focused evaluation passes locally and in CI, results identify failed scenarios and graders, and live benchmarks remain a separate serialized tier.

Written by the indexing model from the issue text.