refactor(core): share the docs guide-read check across docs evals

[czenko]

Problem

build-docs-002-rls-guide builds its guide-read check by hand, and the API keys eval in #212 carries the same function again. Same filter, same hasContent split, same three-way notes string, same comment above it. The only difference is the guide's name.

Ten more docs evals are scoped. Each one points an agent at a page and needs to know whether the agent retrieved it, because without that a red run can't distinguish a guide that taught the wrong thing from a guide the agent never opened.

Solution

• Adds checkDocsGuideRead(toolCalls, { path, label }) to docs-results.ts, beside buildDocsResult. It returns the CheckResult the eval-local copies were building.
• Points build-docs-002-rls-guide at it and deletes the local copy.
• Keeps the check name string byte-identical, so result history doesn't break.
• Covers the three outcomes with tests: guide retrieved, guide reached with no page content, guide not reached.

Manual testing

1. Run pnpm --filter @supabase-evals/core test. 56 pass, including three new checkDocsGuideRead cases.
2. Run pnpm typecheck. Clean. It covers evals/**/EVAL.ts.
3. Read the build-docs-002-rls-guide/EVAL.ts diff. Confirm the name passed to label reproduces the removed check's name exactly.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
evals	Ignored	Preview	Aug 24, 2026 5:10pm

[czenko]

Worth noting, since this is an architectural decision: this is the first CheckResult builder in core. Core has parsed traces into data and left the pass/fail decision to evals until now.

[czenko]

Parking this as a draft per the thread. #212 is unstacked and standalone on main now, carrying its own copy of the check.

Holding until the structured docs-injection idea is decided. If a includeDocs: frontmatter field lands, this check goes vacuous and this PR should just close. If it doesn't, this is still the cleanup once a third docs eval needs it.