test(sessions): consolidate persistence fault scenarios behind an internal harness

[FU-max-boop]

Describe the feature

Recent Session recovery work in #4630, #4650, and #4654 exercises the same persistence boundary through two private list-backed Session doubles:

• tests/test_run_impl_resume_paths.py::_FailingResumeSession
• tests/test_runner_guardrail_resume.py::_ResumeWriteFailureSession

Both need to distinguish an append that raises before mutation from a commit whose acknowledgement is lost. The surrounding regression matrices also repeat the same live/JSON and streaming/non-streaming recovery axes. Keeping those mechanics separate makes it easy for the durability matrix to drift while each individual test still looks reasonable.

I propose a small internal-only fault-injection harness under tests/utils/. It should not be exported from agents.testing.

The harness should provide:

• a list-backed Session with explicit initial history and session ID;
• one-shot add_items() outcomes for success, raise-before-commit, and commit-then-raise;
• detached records of attempted append batches;
• optional support for the existing partial-write or controlled-blocking cases only where current tests already require them; and
• a small immutable case description for run mode and state transport, without encoding expected SDK results.

Runner invocation and caller-visible assertions should remain in the owning test modules. The shared helper should model only Session-boundary mechanics, not become the oracle for recovery behavior.

Acceptance criteria

• Existing recovery coverage remains explicit for:
  • non-streaming and streaming failure/recovery;
  • same-mode and cross-mode recovery;
  • live and JSON-restored RunState;
  • raise-before-commit and commit-then-raise outcomes.
• Exactly-once tool effects, Session history, model input, guardrail results, hooks, and public replay continue to be asserted at the public Runner boundary.
• No runtime package, public import, Session protocol, documentation, or released API contract changes.
• Specialized cancellation, concurrency, compaction, and real-backend tests may keep local doubles when their lifecycle cannot be represented faithfully.
• Coordinate any change to tests/utils/simple_session.py with feat: add approval override argument parity for RunState resumes #2695.

Non-goals

• Adding a public agents.testing.ScriptedSession.
• Claiming atomicity, CAS, distributed exactly-once behavior, or partial-write semantics for real Session backends.
• Changing Session recovery behavior.
• Replacing backend-specific integration tests.

I searched open and closed issues and pull requests for ScriptedSession, scripted Session testing, failing Session helpers, and Session scenario harnesses and did not find a duplicate.

[seratch]

Thanks for mapping these tests. I confirmed that the two private Session doubles overlap in their raise-before-commit and commit-then-raise behavior, but the current coverage has not drifted: test_run_impl_resume_paths.py exercises the full recovery matrix, while the guardrail tests deliberately select representative combinations for their own caller-visible assertions.

Both doubles already build on SimpleListSession, and the duplicated fault behavior is small and explicit. Adding outcome queues, attempt records, and shared case descriptions would introduce a broader test abstraction without fixing a demonstrated missing or inconsistent behavior.

I am going to close this for now. If a concrete matrix omission appears, or a third test needs the same fault semantics, we can extract the smallest common Session double at that point.