Add maintainability sensors to Uncle Bob coding harness

[tninja]

Summary

Follow up on #469 by strengthening the bundled Uncle Bob coding-agent harness with maintainability sensors that cover cross-file and architectural risks, not just file/function-level checks.

This updates both the normal and diagnostics-aware harness prompts.

What changed

• Add repository-native dependency-boundary checks to the gauntlet.
• Add a semantic modularity review for non-trivial cross-file changes, focused on:
  • semantic duplication,
  • misplaced responsibilities,
  • inconsistent patterns,
  • dependency direction,
  • excessive change radius.
• Treat coupling/complexity metrics as risk-triage signals rather than automatic proof of bad design.
• Add explicit guidance for sensor conflicts so the agent does not mechanically optimize one metric while making another design dimension worse.
• Add a final sensor gate immediately before EVIDENCE so completion is based on the final source state, with baseline/trend information when available.
• Extend the evidence report with sensor status and maintainability-review findings.
• Require a second independent modularity pass for high-risk Tier 3 architecture changes, since inferential reviews can surface different issues on separate runs.

Why

The original harness from #469 already covered strong computational feedback such as tests, types, lint, coverage, mutation testing, property testing, real execution, and security/supply-chain checks. The main gap was maintainability across file and module boundaries.

Birgitta Böckeler's article notes that computational sensors work especially well at the file/function level, while modularity and coupling require semantic interpretation. It also highlights that guide-only instructions to check sensors are unreliable, and that final sensor status, baselines/trends, and sensor conflicts are useful parts of a coding harness.

This PR incorporates those ideas without adding backend-specific runtime behavior or a new sidecar subsystem; it keeps the change scoped to the existing harness prompts.

Verification

• Prompt-only change; no runtime Elisp behavior changed.
• Normal and diagnostics-aware prompts were updated in parallel.
• Branch diff contains only the two bundled Uncle Bob harness prompt files.

Source

Inspired by Birgitta Böckeler, “Maintainability sensors for coding agents” (Martin Fowler, May 27, 2026):
https://martinfowler.com/articles/sensors-for-coding-agents.html#ConclusionsAndOpenQuestions

Relevant sections include AI modularity review, Conclusions and open questions, and Integration with the coding harness.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c787230512

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]

Allow unchanged baseline sensor failures

When a repository already has a failing test, lint, type, or architecture sensor, this unconditional blocker prevents completion even if the result is unchanged from baseline, conflicting with the gauntlet's explicit allowance for pre-existing baseline failures. This makes routine work on repositories with known debt require human risk acceptance despite introducing no regression; block only sensors that are newly failing or worse than baseline. The diagnostics-aware copy contains the same issue.

Useful? React with 👍 / 👎.