[Feature]: Exclude frozen Red tests from SWE-bench `model_patch`

[GenTang]

痛点与背景 / Problem and Context

Problem

In the red_green flow, the frozen Red test is materialized into the workspace. SWE-bench evaluation currently collects the complete Git diff, so model_patch contains both:

production fix + generated Red test

This may conflict with SWE-bench’s official test_patch and affect evaluation validity.

期望的解决方案 / Proposed Solution

Keep both deliverables in the workspace, but submit only the production fix to SWE-bench:

Workspace:         production fix + frozen Red test
SWE-bench patch:   production fix only
Trace artifacts:   full patch + Red test patch

Requirements

• Exclude only the exact paths recorded in frozen_test.files.
• Do not broadly filter tests/**.
• Apply filtering only to red_green SWE-bench submissions.
• Keep direct_execute behavior unchanged.
• Fail closed if the frozen test cannot be safely excluded.
• Record excluded paths and submission patch SHA in the trace.
• Add tests covering tracked and newly created Red test files.