feat: Auto-fix mode — apply safe rewrites for H1, H2, H3 findings

[roli-lpci]

Problem

lintlang identifies issues but always requires manual fixes. For well-defined, low-risk findings (vague tool descriptions, missing output format constraints, negative-to-positive rewrites), an auto-fix mode would dramatically improve adoption.

Proposed Solution

lintlang audit AGENTS.md --fix          # Apply safe fixes, show diff
lintlang audit AGENTS.md --fix --dry-run # Show what would be fixed, don't write

Safe auto-fixes (high confidence, low risk):

• H1 (vague tool descriptions): Append Returns: <type> and When to use: <condition> to tool descriptions missing them
• H2 (missing output format): Add output format hint when missing (Format: bullet list | JSON | plain text)
• H3 (style negatives): Rewrite Don't be verbose → Be concise (positive form)
• H3 (missing constraints): Add Scope: <inferred from context> to coding instructions missing scope

Manual-only (too risky to auto-fix):

• Security negatives (keep as-is)
• Priority ordering (requires semantic understanding of intent)
• Cross-file conflicts (can't auto-resolve contradictions)

Implementation Notes

• Always show diff before writing
• --fix writes in place; --fix --backup saves .lintlang.bak first
• Only apply fixes where confidence > 80%

Why This Matters

Auto-fix turns lintlang from a read-only linter into an active improvement tool. Directly competes with Cursor/Copilot for agent config improvement workflows.