[b-re-w]

Refs #149. The concurrent-resize behaviour is still awaiting a user decision, so the issue stays open. numpy is split out to #220.

Feature added:

• array.array[float] and array.array[int] parameters compile without a copy. The entry guard checks the exact type, a C-contiguous 1-D buffer with format d or q/l, itemsize 8 and 8-byte alignment; anything else deopts before any effect.
• The buffer export is held for the call and released on every exit (ok, error, deopt).
• Every element access is one relaxed atomic 8-byte load or store, so each access is atomic like the interpreter's a[i]. There is no copy-in and no critical section, and no limit on the number of buffer parameters.
• Two parameters that are the same array are correct: reads see earlier writes in program order.
• Closedness (TypedPython: an ArrayParam function that calls user code compiles and diverges from CPython #203) applies.
• IndexError text matches CPython (array index out of range).
• -DTP_BUFFER_PLAIN switches element access to plain loads and stores, for measurement only.

Not done:

• numpy, because its scalar types and RuntimeWarnings differ from compiled doubles (verified with numpy 2.5.4). NDArray parameters stay interpreted, and a test pins that their results and np.float64 result type equal the interpreter's. Tracked in TypedPython: numpy array parameters (types, warnings and division follow numpy scalars) #220.
• Open point (TypedPython: zero-copy buffer-protocol arrays (array.array, numpy, memoryview) as ArrayParam #149): while the export is held, another thread's resize of that array raises BufferError. Documented; no test depends on it.

Tests added: test_buffer_arrays.py, 51 tests.

• Single-thread differential against the interpreter: reads, writes, len, IndexError, negative indices, a wrong typecode taking the object path, aliasing, an exception mid-loop, an export released after every exit, closedness, and the plain switch.
• Red before: 74 failed in the first version; the tests that passed then were vacuous until assert_compiled was added.
• Mutation: dropping the release on the error path leaves append raising BufferError, so the test fails.

Measured (quiet window, 11 interleaved pairs, 2,000,000 doubles, 3.15.0rc1t; benchmarks/typedpython/results/2026-10-17-buffer-atomic-*):

kernel	atomic / plain	note
dot (read + reduction)	1.001	the float reduction is not vectorised anyway (strict FP)
scale (element-wise store)	2.539	the atomics block vectorisation of element-wise loops
dot, array.array vs list copy-in	1.94 ms vs 12.28 ms	zero-copy is 6.3x faster

Verification (local, 3.15.0rc1t):

• Full suite as CI runs it: 1138 passed, 3 skipped.
• The compiled function sets of the five benchmarks equal develop's. They use lists, and nbody still needs the pending decision.