Refuse a sync overrun at guard exit even when the timer never fired

[Shashankss1205]

The gap

The sync deadline_guard's exit path raised only when state["fired"] was set. On the threading.Timer + PyThreadState_SetAsyncExc fallback — the mechanism every worker-thread run uses (threaded servers, ThreadPoolExecutor callers) — the timer thread needs the GIL to run fire(). A node that held the GIL through the deadline (a long C-level call, or simply timer-scheduling latency) returned before fire() ever ran; disarm() cancelled the pending timer, nothing raised, and the overrunning node's writes committed. For a last node there is no next _enter budget check, so the run completed past its wall-clock ceiling and reported success — violating the README's "a node that overran does not get its writes into state" and the guard's own docstring. The async guard already refused this exact case at the same boundary (grapharc/runtime/graph.py:255-260).

The fix

• grapharc/runtime/budget.py: mirror the async guard's exit check — after disarm(), raise NodeDeadlineExceeded(detail()) when state["fired"] or meter.remaining_seconds() is not None and <= 0, with a comment naming the timer-never-fired case.
• Docstring: the C-call caveat now reads "is not interrupted mid-call, but the guard still raises on exit", and the closing promise matches the async guard's ("Short of the never-returns case…"), covering the no-interrupt-delivered path.
• tests/test_budget_enforcement.py: test_an_overrun_is_refused_at_exit_even_if_the_timer_never_fired runs deadline_guard on a worker thread with threading.Timer monkeypatched to a timer that never fires and a body that outlasts max_seconds, asserting the guard raises on exit. This tests the exit contract deterministically rather than racing the timer thread (per the cautionary notes around _run_swallower). It fails on unfixed code (guard exits cleanly) and passes with the fix.

Out of scope, untouched per the issue: the never-returns case, the SIGALRM mechanism, and _async_deadline.

The issue's repro (switch-interval GIL hold under Budget(max_seconds=0.2) driven from a worker thread) now prints {'exc': NodeDeadlineExceeded("max_seconds reached while node 'slow' was running (0.6s/0.2s)")} instead of a committed result. README's "Budgets" paragraph is accurate again as written, so it is unchanged.

Verification

• pytest: 1648 passed, 12 deselected (full suite, no flaky reruns needed on this run)
• ruff check .: clean
• Existing swallower / re-arm / refuses-to-start tests pass unchanged

Fixes #22

🤖 Generated with Claude Code