[b-re-w]

Closes #141 (phase 1; phase 2 is folded into #159; the open decision is #152).

Defect fixed: compiled loops and recursion never ran the eval breaker. KeyboardInterrupt did not reach a compiled loop, other threads starved, and on free-threaded builds a stop-the-world request would wait for the loop to end.

Feature added:

• Every loop iteration and every impl entry polls. Loops use a register countdown (tp_pc_tick), entries the per-thread counter.
• Every TP_POLL_INTERVAL polls the out-of-line, cold tp_poll_slow calls a module-held lambda: None. Its RESUME runs CPython's own eval-breaker handling: signals, pending calls, forced GIL switching, async exceptions and scheduled GC, exactly as between bytecodes. Public API only.
• A function holding an array copy-in or an entry-globals snapshot never polls, and defers polls while it is active, because the user code run by a poll could change what it copied.

Tests added: test_eval_breaker.py.

• Red before the fix: 3 of 4 failed (interrupt in a loop, interrupt in recursion, a second thread runs). The deferral test passed already.
• Mutation: with the deferral removed, the handler runs mid-call and its write is lost, so the test fails.

Measured cost (quiet window, load 2 to 4, interleaved against develop, 11 pairs; raw data in benchmarks/typedpython/results/2026-10-17-eval-breaker-*.json):

	after / before, median	range
fannkuch-redux	1.082 (+8.2%)	1.043 to 1.191
fannkuch-redux, loop polls removed (experiment, not shipped)	1.000	0.994 to 1.099
nbody	0.995	0.882 to 1.043

The whole remaining cost is the per-iteration countdown in short inner loops; CPython pays the same kind of check on every back-edge. Accepted by the leader as the price of the fix. Chunked polling for range loops is #206.

Verification (local): full TypedPython suite on CPython 3.14.7: 966 passed, 3 skipped.

[b-re-w]

Re-measured after a report that another session ran a model test for about 60 s while the MEASURING flag was set earlier today. Both re-runs started in a quiet window (5-minute load at most 3, no torchnative gate running), interleaved, 11 pairs:

	first run (16:29 / 16:32)	re-run (16:42 / 16:45)
fannkuch-redux, after / before	1.082 (1.043 to 1.191)	1.080 (0.835 to 1.138)
nbody, after / before	0.995	1.001
fannkuch-redux, loop polls removed	1.000	1.007

The figures agree, so the recorded numbers stand.