agent: clean up the proof tree files a run leaves behind

[dingf3ng]

Why

ProofController writes <theorem>_proof_tree_final.png and .json next to the file being proved (proof_controller.py:280).

They're named after the theorem, not the scratch copy — so ScratchProof's cleanup, which works from self.path, never matched them. Five benchmark goals left ten files in AutoRocq-bench, untracked and not gitignored.

That undercuts the whole promise of the scratch copy: your tree comes back untouched.

What changed

Deleting by name pattern would risk removing output from an earlier run. So instead:

• open() records which *_proof_tree_final.* already exist
• close() removes only the ones this run added

Same shape as the _CoqProject backup already in the class.

Result

Same five goals re-run. Afterwards the benchmark tree is empty on both checks:

git status --porcelain          (nothing)
*_autorocq_*.v                  (nothing)
*_proof_tree_final.*            (nothing)

[dingf3ng]

Superseded by #17.

Two things were wrong here:

1. The description claimed the files were "untracked and not gitignored" as a general fact. The root .gitignore scopes *.png / *.json to proof-search/examples/ (lines 36-37), so runs against the examples were always covered. Only AutoRocq-bench — a submodule with no such entries — was affected.

2. Deleting the files was the wrong repair. The proof tree is a run artifact in the wrong directory, not litter: ProofRecorder writes statistics and never the tree, so those two files were the only copy of it. This PR removed the output instead of relocating it.

#17 writes the tree into the run's output directory instead, beside autorocq.log and the resulting .v, and leaves utils/scratch.py untouched.