Require compilable GenPlan code submissions

[joshnroy]

Summary

Stacked on #177.

Treat each GenPlan implementation response as an automated code submission:

• tell the model initially and on every repair that its response is compiled as Python, not read as prose;
• allow commentary only as Python comments inside the single complete module;
• compile extracted source before promotion;
• preserve the previous policy when a submission does not compile;
• return compiler feedback to the same budgeted refinement loop.

A rejected submission still counts as a model generation and toward the dollar budget. The checker only compiles; ordinary loading and training validation continue to catch missing interfaces and runtime failures.

Verification

• Focused offline suite: 42 passed, 2 paid integration tests deselected.
• Four new contract/checker/preservation tests: passed.
• Black, isort, diff check, targeted pylint (10.00/10), and targeted mypy passed.

Experiment isolation

Implemented and verified in an isolated worktree. The active sweep in the main workspace was not modified or restarted.