Chunk's dynamic-sizing ceiling was discarding real GPU headroom

dense_evolution.chunk.get_dynamic_chunk already reads the ACTIVE compute device's real available memory (device.memory_stats() on GPU/TPU, falling back to psutil.virtual_memory() on CPU) -- that fix shipped earlier. But the function's own final clamp, max(16, min(max_bits, 27)), threw away real headroom regardless of what the device actually reported: 27 bits is only 2.1 GB, so any GPU with more than ~2.5 GB free (nearly every real GPU, Kaggle's T4/P100 included) was being capped well below what device.memory_stats() itself said was safely usable.

What was verified, not assumed

Ran a real Kaggle GPU kernel (T4/P100) that imports dense_evolution directly and calls get_dynamic_chunk(jnp.complex128):

• Before (existing PyPI code, ceiling at 27): would have printed 27 regardless of the device's real free memory.
• After (ceiling raised to 30 in a local branch, not yet promoted): printed 29 on the real GPU -- confirming the device-memory read was already correctly computing a value above 27, and the old hardcoded ceiling was the only thing suppressing it.

Why 30, not unbounded

30 bits (~17 GB per chunk) is still a real ceiling -- avoids ever computing a chunk_size_bits large enough to risk int32-indexing issues elsewhere in the codebase -- it just no longer discards headroom a modern GPU actually has. The function's own memory-based calculation already caps the actual returned value at whatever the device really supports; raising the ceiling only stops it from being overridden below that on hardware with more VRAM than a 27-bit chunk needs.

Status

Documented here first, per this repo's promotion convention -- not yet merged into Dense-Evolution (the closed PR there, #278, explicitly deferred to this write-up). Next step: re-propose the same fix as a new PR on Dense-Evolution referencing this verification.