1.Why is 'handles simple, well-specified bug-fix issues in Python repos under 10k LOC' a strong scope statement for the capstone?
2.Why is codebase exploration fundamentally a retrieval problem for this capstone?
3.For a coding agent locating a specific bug, why does agentic grep-and-read often beat pure embedding-based retrieval?
4.What is the main advantage of search/replace edits over full-file rewrites for bug fixes?
5.Why must the repair loop write a test that reproduces the bug and fails BEFORE the fix?
6.Why must the test-driven repair loop be bounded (e.g., max 5 attempts)?
7.When the model says it's finished fixing the bug, what should the loop do?
8.Why is opening a PR treated as an irreversible action requiring an HITL gate?
9.What should a self-assembled SWE-bench-style eval set contain, and why?
10.Why report a partial-success taxonomy instead of just a pass/fail rate?
11.What makes a limitations doc the artifact that most signals seniority?
12.In a system-design interview, which move most reliably signals seniority early?