TL;DR

Two small fixes to make coding-loop execution deterministic across builds:

• Fix 1: permission_mode is now propagated from build config into coder/reviewer/QA/synthesizer calls.
• Fix 2: iteration checkpoint files are now scoped by build_id to prevent stale resume state leaking between builds.

Included tests

• test_run_coding_loop_propagates_permission_mode_to_all_agents
• test_run_coding_loop_ignores_legacy_iteration_state_when_build_id_present

Validation

• AGENTFIELD_SERVER=http://localhost:9999 pytest -q tests/test_coding_loop_regressions.py (2 passed)
• ruff check on touched files (clean)

Summary

This PR fixes two execution-loop issues that can cause inconsistent behavior across builds:

1. permission_mode from build config was not being propagated into coding-loop agent calls.
2. Iteration checkpoints were keyed only by issue name, so stale state from prior builds could be resumed unintentionally.

Changes

• Propagate permission_mode through execution config plumbing:
  • Include permission_mode in BuildConfig.to_execution_config_dict().
  • Use config.permission_mode in coder/reviewer/QA/synthesizer calls.
• Scope iteration checkpoints by build:
  • Checkpoint path now supports execution/iterations/<build_id>/<issue>.json.
  • Load/save checkpoints using dag_state.build_id.
• Add regression tests for permission-mode propagation and stale-state isolation.

Why this is safe

• Existing behavior is preserved when no build_id is available (fallback path still exists).
• Resume behavior for the same build is unchanged.
• Changes are isolated to execution config plumbing + coding loop checkpoint pathing.