ci: ARM oracle job + integer-kernel bit-exactness gate (closes #1081)

[JustVugg]

Closes #1081.

What

1. tests/test_int_kernel_exact.c — integer kernels have no rounding excuse: against a pure-C reference on the same inputs they must match bit for bit, on every ISA. Covers dot_i4i8, the planarize_i4 round-trip, dot_i4p_u (K1 plane layout, the −8·sum(x) identity), and matmul_i4p_idot vs matmul_i4_idot bitwise. 2,176 checks, 0 failures on x86 (avx-vnni branch live).
2. oracle-arm job (ubuntu-24.04-arm, timeout-bounded per ci: stop running every dev commit twice, and bound every job #953's direction): builds every engine — NEON compile coverage; K1's dot_i4p_u NEON branch had never been compiled by anyone — runs the exactness gate with the NEON branches live, and replays the glm_tiny teacher-forcing oracle against a fixture generated on the same runner (same-machine torch reference, no cross-ISA float excuses).
3. The x86 efficiency job now runs the same exactness gate, so both ISAs hold the same bar.

The test found a real contract on day one

The first draft generated full-range int8 activations and produced 30 false failures against the shipped dot_i4i8: x = −128. The abs/sign VNNI trick cannot represent −(−128), and qrow_i8 clamps to [−127,127] — so −128 never reaches the kernels in production. The test now encodes that contract explicitly, with a comment: it is the kind of invariant that only exists in people's heads until a test writes it down.

Deliberately deferred

deepseek-v4-tiny-check on ARM — its ARCH/PORTABLE_ARCH plumbing needs checking for aarch64 first. qwen36's tiny gate joins when #712 lands (its author's review is what surfaced this class — credit @kreuzzelg).

⚠️ Note for review: this PR is also the first time K1's NEON branch compiles anywhere. If the ARM job fails on it, that is the job doing its work — the fix lands here.

🤖 Generated with Claude Code