ktext init: scan CLAUDE.md, AGENTS.md, and other agent instruction files

[thegagne]

Problem

ktext init currently scans README, Makefile, package manifests, ADRs, directory structure, and CONTRIBUTING.md — but not agent instruction files like CLAUDE.md, AGENTS.md, .cursorrules, or copilot-instructions.md.

These files often contain exactly the kind of high-signal content that belongs in CONTEXT.yaml: constraints, conventions, architectural decisions. Ignoring them means users who already have them have to re-enter that information manually.

This also directly addresses the common objection "I already have a CLAUDE.md" — ktext doesn't replace it, it absorbs it, structures it, scores it, and makes it portable across tools.

Proposed behaviour

During ktext init, the scanner should:

1. Detect the presence of any of these files:

   • CLAUDE.md
   • AGENTS.md
   • .cursorrules
   • .github/copilot-instructions.md
   • .cursor/rules
   • .windsurfrules
   • .clinerules

2. Read their content and attempt to extract candidates for:

   • constraints — lines containing must, never, always, avoid, prohibit, etc.
   • conventions — lines with action verbs (use, run, write, call, etc.)
   • decisions — sections that explain why something was chosen

3. Surface extracted candidates in the interactive review — same flow as other discovered values, user accepts or rewrites each one

Test case

The Backstage repo has an AGENTS.md file. That would be a good real-world test of whether the extraction produces useful candidates vs noise.

Notes

• Extraction doesn't need to be perfect — the interactive review is the safety net
• Should log which files were found during the "Scanning repo..." output
• Low-quality or very short instruction files should still be processed — the scorer will penalize vague entries