Make arena alignment absolute via a base pad

[AntonOresten]

Arena: pad the origin so alignment is absolute

Arena's alignment was defined relative to the slab's first byte, with absolute alignment documented as the slab provider's problem: fine for CUDA's 256-aligned bases, silently short for a Julia Vector's 16/64-aligned base, and an actual correctness fault for odd-based view slabs (a Float32 carve at an absolutely-misaligned address is a kernel fault on a GPU, not a perf note).

The arena now pads its origin by -base mod alignment at construction, when the storage exposes a base address through the new slabbase(slab) hook:

• DenseVector{UInt8} answers UInt(pointer(slab)) — covers Vector, Mmap, CuArray, JLArray with no extension.
• The generic answer is nothing: no pad, today's relative contract, for storages without a stable pointer.

Carve addresses are then divisible by alignment wherever the slab landed, costing at most alignment - 1 bytes, once. For CUDA slabs the base is already 256-aligned, so the pad is zero and nothing changes. Offsets, marks, and the watermark stay relative to the padded origin — mark/retract!/reset! and address determinism under capture are untouched. The watermark docstring now notes the + alignment slack when sizing a fresh slab, since a differently-based slab pads differently.

The base is a virtual address, which is the correct notion here: every alignment consumer (cuBLASLt/cuDNN contracts, load/store units) checks the VA, and pages map whole so VA ≡ PA modulo the page size — far above any alignment an arena is asked for. The one real hazard, a base that can move, is made part of the slabbase contract: such storage must answer nothing.

CaptureViolation: docstring rewritten around remeasured ground truth

Re-verified the failure modes on driver 13.3 / CUDA.jl 6.2.1. The old docstring claimed freeing inside the captured region "does not rescue it: the free path errors during capture" — that is no longer true. The rewrite splits the hazard as it actually behaves:

• Free outside the region or absent (no-op release!, frame closing after capture, escaped buffer): unpaired allocation node, graph launches once, second launch fails with ERROR_INVALID_VALUE, GC-timing dependent.
• Free inside the region: nodes pair, replay is legal and measured correct — but it stands on driver behavior CUDA.jl documents as unsupported, pays per-replay map/unmap, records fresh addresses on re-capture, and a pool-exhausted allocation mid-capture still escalates to device_synchronize, invalidating the capture with GC disabled.

The guard's doctrine is unchanged — GC spaces refuse under capture, arenas carve — the justification is just no longer stale.

Tests

Two slab-aliasing tests index above the pad; a new testset covers carve-address divisibility on a Vector slab, the nothing-base fallback for view slabs, and the pad counting against the ceiling. 181/181 pass, plus a live check on a CuVector slab (pad 0, carves 256-aligned).

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!