Harden Uncle Bob harness against false-green checks

[tninja]

Summary

Follow up on #469 by selectively incorporating the highest-value hardening from the current AmazingAng/old-coder skill without turning the bundled harness into a mirror of upstream.

This PR focuses on failure modes where the gauntlet can look green without actually enforcing the intended constraint.

Changes

• require every gauntlet layer to be an executable gate, not merely a report
• require coverage gates to fail non-zero when the requirement is missed
• prefer established mutation tools; manual mutation must prove each mutant was applied and executed
• make custom/home-grown checkers fail closed and require a known-bad negative control before trusting green
• distinguish N/A, UNAVAILABLE, and SUBSTITUTED layers in EVIDENCE
• require orchestration to verify that every declared required layer actually ran
• clarify that answering a spec clarification question is not approval of the revised spec
• persist the approved spec when practical and tighten isolated-working-tree guidance

Both bundled prompt variants are kept in sync; the diagnostics variant retains its existing diagnostics baseline/get_diagnostics constraint.

Why

Recent upstream old-coder dogfooding found several important false-green cases: coverage commands that only printed percentages, manual mutation runners that could report kills without executing the intended mutant, and gauntlet orchestration that could stay green when a layer never ran. These are general harness lessons worth carrying into ai-code-interface.el.

I intentionally did not vendor upstream's independent verifier protocol, old-coder-api, demo-specific source-state machinery, or other ecosystem-specific additions. The goal is to cherry-pick durable harness principles rather than synchronize prompt text.

Upstream reference point reviewed: AmazingAng/old-coder through commit a0eb529d393a1cb3ccc564e32b2104e7e75c7a29 and the current skills/old-coder/SKILL.md.

Testing

Prompt-only change; no Elisp runtime behavior is modified. The two prompt variants were compared together to keep their shared harness contract aligned.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 32d87685da

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Apply the hardening to the Harness+ prompts

When a user selects uncle-bob-coding-agent-harness-plus, ai-code--auto-test-harness-file-name routes them to the separate Harness+ or Harness+-diagnostics file, but both still contain the old report-only coverage and mutation guidance and omit the new fail-closed checks. Consequently, the selectable maintainability-enhanced harness remains vulnerable to the exact false-green behavior this change is intended to prevent; update both Harness+ prompt variants with the shared hardening clauses as well.

Useful? React with 👍 / 👎.