dotnet-aot-compat: evaluation scenario too large to complete within timeout

The `dotnet-aot-compat` evaluation scenario (Azure.ResourceManager — 461 files, 135+ JSON serialization call sites) consistently hits the 1050s timeout in both baseline and withSkill runs. This means:

1. **We can't measure skill quality** — every run times out before the agent finishes, so judge scores reflect "how far did it get" rather than "did the approach work." Per-run variance is huge (improvement scores ranged from -0.20 to +0.22 across runs of the same wording).

2. **Real users likely hit the same wall** — a user with Copilot working on a similarly large codebase will get an incomplete result regardless of how good the skill guidance is.

### Proposed follow-ups

#### 1. Use a smaller test project for evaluation

Replace or supplement Azure.ResourceManager with a project that has ~20-30 AOT warning sites instead of 135+. This lets the agent actually complete the task within the timeout, making skill improvements measurable. The current scenario is useful as a stress test but not as a quality signal.

#### 2. Add scope-limiting guidance to the skill

When the codebase has a very large number of warnings, the skill should guide the agent to prioritize rather than attempt to fix everything. For example:
- "If there are more than ~30 distinct warning sites, focus on the highest-impact patterns first (e.g., JSON serialization, reflection-based factories) and document remaining warnings as TODOs"
- This helps both evaluation (agent can finish in time) and real users (they get a useful partial result instead of a timed-out incomplete one)

### Context

This was discovered during analysis of #366 evaluation results. See the [detailed comparison comment](https://github.com/dotnet/skills/pull/366#issuecomment-4059507897). The batching guidance in #366 is a step in the right direction (judge confirmed the agent actually attempted builds with new wording vs. zero builds with old), but the timeout is the dominant factor masking any improvement.