discovery(tests): orchestration suite has 5 pre-existing failures — checkpoint test rot (_checkpoint_manager rename) + workflow_memory isolation pollution #11754

Discovered during

Implementation of #6828 (AgentRegistry consolidation, 2026-07-15) — surfaced when running the full autobot-backend/orchestration/ suite as a pre-commit verification gate. Reproduced identically on a clean Dev_new_gui tip (7a95b99), so both are pre-existing and unrelated to the #6828 diff.

Finding 1 — workflow_executor_checkpoint_test.py: 3 deterministic failures (test rot)

autobot-backend/orchestration/workflow_executor_checkpoint_test.py:107:
    executor._checkpoint_manager.save = save_spy
E   AttributeError: 'WorkflowExecutor' object has no attribute '_checkpoint_manager'. Did you mean: '_checkpoint_resumer'?

Failing tests (always, even in isolation):

• test_checkpoint_failure_does_not_propagate_in_execute_step_with_agent
• test_checkpoint_success_called_once_on_happy_path
• test_checkpoint_not_called_when_step_fails

The production attribute was renamed _checkpoint_manager → _checkpoint_resumer without updating the test's spy target. The tests need re-pointing at _checkpoint_resumer (and re-verifying they still assert the intended checkpoint semantics rather than just the attribute name).

Finding 2 — workflow_memory_test.py: 2 ordering-dependent failures (test isolation)

• TestWorkflowMemoryAutoInjection::test_prior_findings_injected_into_context
• TestWorkflowMemoryAutoInjection::test_empty_memory_not_injected

Pass when the file runs alone (34 passed), fail when run as part of pytest autobot-backend/tests/orchestration/ autobot-backend/orchestration/ ... — cross-test state pollution (likely a shared singleton/module-level cache mutated by an earlier test). Needs a fixture-level reset of whatever seam the auto-injection reads.

Evidence

# worktree AND clean base tip both:
python3 -m pytest autobot-backend/orchestration/workflow_executor_checkpoint_test.py -q
# -> 3 failed
python3 -m pytest autobot-backend/orchestration/workflow_memory_test.py -q
# -> 34 passed (alone)
python3 -m pytest autobot-backend/tests/orchestration/ autobot-backend/orchestration/ autobot-backend/services/mesh_brain/ -q
# -> 5 failed, 708 passed

Why it matters

Deterministic red tests in the orchestration suite mask real regressions in checkpointing, and ordering-dependent failures make full-suite runs untrustworthy as a merge gate.