Eval: llms.txt vs raw HTML as doc source

[sirugh]

Problem

The per-source guides in skills/docs/ currently point agents to llms.txt bundles as the primary fetch target. But it's not validated whether agents perform better when they fetch the llms.txt-bundled content vs. fetching the underlying HTML pages directly.

Hypothesis

llms.txt bundles may be denser and more token-efficient, but HTML pages may contain additional structured content (code blocks, nav, metadata) that helps agents answer accurately. Or vice versa — the bundle may already strip noise and give better answers.

What to test

Using the existing claude-ext eval harness (scripts/run-evals.py):

1. Baseline: current behavior (fetch llms.txt bundles)
2. Variant: modify a per-source guide (e.g. storefront.md) to instruct the agent to fetch the HTML source URL instead of the llms.txt bundle for the same topic
3. Run the same eval set (e.g. commerce-storefront.json) against both variants with --runs 3
4. Score with scripts/score-evals.py and compare pass rates

Acceptance criteria

• At least one domain (storefront or backend) is tested with 3+ runs per variant
• Results are saved to results/ with a clear naming convention distinguishing the two variants
• A short summary in this issue or a doc update with the recommendation: stick with llms.txt, switch to HTML, or adopt a fallback pattern