docs: update mkdocs content for yeti

[frostyardyeti]

Summary

Updates the issue-refiner job documentation to reflect recent prompt improvements that make implementation plans more rigorous and grounded in actual code. These changes document stronger guardrails around phantom references, scope discipline, step-by-step buildability, and a fifth self-critique dimension (completeness vs. gold-plating).

Changes

• Phantom detection: Documented that Claude now flags functions, types, APIs, or file paths that don't exist in the codebase rather than planning around them
• File path verification: Plans must only reference files that were actually opened and confirmed to exist
• Step buildability: Each implementation step must leave the project in a buildable, test-passing state
• Expanded risk coverage: Risks section now explicitly includes concurrency, error paths, and boundary conditions
• Testing patterns: Plans must follow the repo's existing test framework, mock style, and fixture conventions
• Scope guardrails: Added explicit "What NOT to plan" guidance — no unnecessary logging, README updates, impossible-scenario validation, or adjacent code cleanup
• Fifth self-critique dimension: Added "completeness vs. gold-plating" to the structured self-critique, ensuring plans solve the full issue without over-solving