[b-re-w]

Closes #152 (decision (i), user-approved: CPython identity).

• SPEC N-8: a function holding a list copy-in or an entry-globals snapshot now polls inside long loops, from the second TP_POLL_INTERVAL chunk on, and each While countdown. At each poll:
  1. Write the lists back, as CPython's own stores would already be visible.
  2. End the critical section, poll, and take it again.
  3. Copy the lists in again and re-check lengths, exact element types and entry globals.
  • Unchanged: it continues compiled.
  • Changed: an OSR-style exit to a generated interpreted continuation of that iteration. The call returns or raises what the continuation does.
• resume.py (frontend) generates one continuation per loop from the original AST:
  • The innermost loop continues from its current index.
  • Each enclosing loop runs the rest of its current iteration (a first-iteration flag keeps break/continue exact) and then its remaining iterations.
  • Range bounds are the values evaluated at loop entry.
  • The original lines and co_name/co_qualname are kept, so tracebacks match.
  • Unbound locals are passed as a sentinel and deleted.
• cgen and runtime:
  • The wrapper passes an OSR context (lock, lists, result) through tp_tls. The impl takes it at entry, so internal and recursive calls never poll with someone else's lists.
  • At an exit the impl boxes every live local (a local array becomes a new list) and calls the continuation. The wrapper skips write-back after rc 2, because the lists are already final.
  • Functions with virtual objects or regions, or a parameter that cannot be handed over, keep the TypedPython: compiled code never checks the eval breaker (signals, GC, GIL hand-off) #141 no-poll rule.
• Short loops (under the interval) never poll, so a short call stays one critical section, still an execution CPython can produce.
• Tests:
  • test_resume.py (33): 320 resume points, each compared with the interpreted run.
  • test_osr_exit.py (19): deterministic, with TP_POLL_INTERVAL=32 and the test-only TP_TEST_SNAP_HOOK running Python in the poll window. Covers unchanged (no exit), length change, element type change, entry-global rebinding, nested loops with break/continue, while, an exception in the continuation, and an exception at the poll with CPython's partial writes.
  • Mutant: never taking the exit crashes the process on the length-change cases. Bounds proven against the old len no longer hold, so the re-check is also memory safety.
  • test_free_threaded: the two whole-call tests now use lists below the interval and explain why. Longer loops are visible chunk by chunk, as interpreted stores are. 25/25.
  • test_regions: the allocation check is one-sided (a call can free other tests' garbage in the full run).
  • Full TypedPython suite locally: 1678 passed. The one failure was that allocation count, now fixed. 27/27.
  • Benchmark compiled-function sets are identical to develop (7 files).
• Not covered: OSR for ForObj sites (they never occur in a snapshot holder, which must be closed), and functions with virtual objects or regions (no poll, as before).