Cross-repo evaluation harness

[JSONbored]

Problem: Without a real benchmark, there's no way to know whether the "generalized" agent is actually general, or just quietly still tuned to gittensory's own repo shape.

Area: AMS / Miner / QA

Proposal: Assemble a benchmark set of real, diverse external repositories (different languages and stacks) and run the full loop against each, scoring success objectively (task completed, target repo's own CI passing, no gittensory-specific assumption triggered).

Deliverables:

• A benchmark repo set and a repeatable evaluation run producing a pass/fail report per repo.

Acceptance criteria:

• The loop succeeds against a majority of the benchmark set without any gittensory-specific configuration being required.
• Failures are categorized (stack-detection gap vs. execution gap vs. something else) rather than left as an unexplained pass/fail.

Boundaries:

• Blocked on De-hardcode discovery from gittensory's own conventions #4784 through Generalized plan → implement → test → submit loop #4786.

---

Part of #4778.

[linear-code]

JSO-3460