cuda: direct-I/O fallback observability -- counters, always-on disable log, report line

[nexus-cw]

Addresses #687.

An EINVAL/EFAULT/ENOTSUP/EOPNOTSUPP on the O_DIRECT model read path permanently disables direct I/O for the process with only a DS4_CUDA_WEIGHT_CACHE_VERBOSE-gated stderr line as evidence -- the streaming fast path can silently degrade to buffered I/O forever. This makes the state and history observable, with no behavior change on the I/O path itself:

• Counters (ds4_cuda.cu): fallback events by errno class (einval / efault / enotsup / other-transient), engaged/disabled/unavailable state + the disabling errno, and widened-read stats (count + total wasted alignment bytes). Plain non-atomic counters on the single CUDA-dispatch thread, same convention as the surrounding code; readers may see a beat-stale value, fine for monitoring.
• Always-on log on the permanent-disable event (a permanent behavior change should not be verbose-only). The engaged/unavailable logs stay verbose-gated as before.
• Surfaces: ds4_gpu.h accessors (ds4_gpu_direct_io_state / _disable_errno / _counters) and a self-report line appended to the existing --memory-report output.

Evidence from a GB10 streaming deployment (DeepSeek V4 Flash, --ssd-streaming) running this instrumentation:

ds4: CUDA direct I/O: state=engaged disable_errno=0 fallbacks einval=0
efault=0 enotsup=0 other=0 widened_reads=69166 widen_wasted=270.179 MiB
avg_waste=4096.0 B/read

i.e. exactly one 4 KB alignment page of waste per widened read (~0.03% of a typical expert fetch) -- and a silent downgrade to buffered I/O now reports itself.

Build: make cpu clean (x86_64), ds4_cuda.o + ds4.o compile clean with nvcc (arm64 GB10). Companion proposal: #688 / its PR.

🤖 Generated with Claude Code