[Uyuz-ux]

Fixes #103.

The bug

In the Linux implementation of randomBytes() the retry loop for partial getrandom(2) reads computed the resume pointer p, but then passed the original buffer start pbytes to the syscall:

while res < nbytes:
  p = cast[pointer](cast[uint](pbytes) + uint(res))
  let bytesRead = syscall(SYS_getrandom, pbytes, nbytes - res, 0)  # should be `p`

After a partial read of k bytes, every retry restarted at the buffer start while res kept accumulating, so the region [k, nbytes) was never written and the function still returned full success. For a CSPRNG API that means keys/nonces/salts can silently contain uninitialized memory — exactly as described in #103.

The fix

One line: pass p instead of pbytes, matching how the other implementations in this file (urandomRead, the OpenBSD path) already handle the resume pointer.

Verification

I could not force a partial read deterministically from userspace (this kernel fills large requests in a single call), so I verified the resume path with a small LD_PRELOAD harness that intercepts libc syscall() and makes the first getrandom call with len > 8 return a partial read of 8 bytes, passing later calls through unchanged. A checker program fills a 256-byte buffer with the 0xAB marker, calls randomBytes(), and reports how many stale marker bytes remain after the first 8 positions:

=== upstream master (before fix) ===
without harness: returned=256 staleTailBytes=0 (no partial read -> passes)
with harness:    returned=256 staleTailBytes=248
                 RESULT: FAIL - buffer tail still holds stale memory

=== with this fix ===
with harness:    returned=256 staleTailBytes=0
                 RESULT: PASS - whole buffer filled with random data

So with the old code the tail of the buffer (248 of 256 bytes) kept its pre-call contents while the function reported success; with the fix the whole buffer is filled.

tests/testsysrand.nim passes in all four compile modes used by nimble test (default, -d:danger, --threads:on, --mm:orc --threads:on) on Nim 2.2.4 / Linux amd64.

I did not add a regression test because a partial getrandom(2) read can't be triggered deterministically from a normal unit test (it depends on signal timing / kernel limits), and mocking would require refactoring the direct syscall import. Happy to add one if you'd prefer a refactor that injects the syscall — happy to follow up in a separate PR.