test: strengthen eval slice — realistic cases or explicit scaffold framing

[constk]

Problem

The eval slice is the thinnest part of the harness:

• eval/golden_qa.json contains a single toy case (echo-hello).
• The nightly eval workflow is disabled by default.
• docs/EVAL_HARNESS.md documents the protocol but doesn't frame the "why would I add a case?" question.

A reader expecting an LLM-eval story finds infrastructure without conviction.

Proposed solution

Pick one of the following:

Option A — Add real cases. Add 3–4 realistic golden cases (e.g. a Q/A pair, a code-generation prompt with a deterministic expected substring, a tool-use trace check). Wire the nightly workflow to run them by default. Update EVAL_HARNESS.md with a "how to add a case" section.

Option B — Reframe as scaffold. Add a paragraph at the top of EVAL_HARNESS.md explicitly framing the eval slice as "scaffold for your project's eval cases, not a benchmark." Document the contract for adding cases. Keep the toy case as a smoke test and explain why the nightly is opt-in.

Acceptance criteria

• One of Option A or Option B chosen and implemented
• EVAL_HARNESS.md tells a reader either how to add cases or why this is a scaffold
• Nightly workflow status (enabled / opt-in) is consistent with the chosen framing

Priority rationale

Currently the weakest organ of the harness story; either path closes the gap.

[constk]

Closed via #104 (merged to develop as commit 18b4d30). Awaiting the next release: PR to reach main.