fix: pin SDK to known-good 500-image build SHA

[simonrosenberg]

Summary

• pin the vendored software-agent-sdk submodule to 30819566b56b592cedcd1dc0d503a454f8ca6668
• restore the last SDK revision already verified by successful end-to-end image builds for this regression
• keep the benchmark-side code path unchanged so the fix is isolated to the proven regression surface

Why this fix

Issue #504 tracks the 500-image build regression that followed the SDK bump bundled into #456. Recent repo history shows:

• successful full SWE-Bench build: run 22967520137 on SDK 30819566b56b592cedcd1dc0d503a454f8ca6668
• successful full SWT-Bench build: run 22967516108 on the same SDK SHA
• later SDK SHAs regress again into much slower or cancelled runs

This branch pins benchmarks back to the last SDK SHA already proven to complete both workflows end to end.

Validation

• make build
• uv run pytest tests/test_llm_config.py
• dispatched full branch workflows:
  • SWE-Bench: 22985173176
  • SWT-Bench: 22985173172

Fixes #504

[all-hands-bot]

Approach is correct (rollback on regression), but the PR description has unfilled template placeholders where evidence should be. See detailed feedback below.

🟡 Verdict: The submodule rollback itself is pragmatically sound, but the PR description needs completion before merge. The claims about "known-good SHA" and workflow run links are blank/missing - fill these in to substantiate the rollback justification.

[simonrosenberg]

Pinning the sdk to an old version is not an acceptable fix since it won't allow to evaluate new sdk versions.