feat(sdk): resume an existing run after a crash

[prashar32]

Why

The crash-resume moat is fully supported server-side — on restart the daemon reloads non-terminal runs with their already-spent budget, so a SIGKILL’d run keeps enforcing without re-spending. But the Python SDK could only create a run (governed_run); it had no way to attach to an existing one, so the resume story wasn’t reachable from Python without dropping to the low-level client.

What

Runtime.resume_run(run_id) — a context manager that attaches to an existing run by id. Unlike governed_run it neither creates a new run nor cancels on error: the run keeps its server-side budget and spent usage, so a resumed agent fetches its last checkpoint and continues where it left off, and can’t overspend by restarting.

with rt.resume_run(run_id) as run:
    cp = run.latest_checkpoint()
    start = cp["payload"]["cursor"] if cp else 0
    for i in range(start, total):       # skip the steps you already paid for
        run.step()                      # counts against the SAME budget
        run.checkpoint("step", {"cursor": i + 1})

Verified the server side first (the Day-8 spike)

Before writing the SDK, I proved the protocol against a real kill -9: a run advanced 5 of 8 loop steps → kill -9 the daemon → restart (resumed runs from store count=1, loops still 5, checkpoint intact) → continued: steps 6/7/8 → 200, step 9 → 402. Exactly 3 more allowed (8 total across the SIGKILL), not 8 more — no re-spend.

Tests + docs

• test_resume_run_attaches_without_creating — attaches by id, sets current_run, reads the pre-crash checkpoint, keeps stepping the same run. The stub daemon now serves GET /v1/runs/{id}.
• SDK README gains a Resume after a crash section; CHANGELOG updated.
• Full SDK suite: 16 pass (2 framework-integration tests skip without langchain/openai-agents).