Fix out-of-range integer brng reading past brng_list

[vchamarthi]

Issue

brng_list has BRNG_KINDS = 11 entries, irk_brng_t defines 0–10. The integer
path of _parse_brng_token_ coerced the user value straight to irk_brng_t, which
only checks it fits the enum's underlying integer type. The seeding routines then
index brng_list[brng].

brng=11 returned uninitialized memory, no warning, no error:

mkl_random.MKLRandomState(1, brng=11).randint(0, 100, 6)
array([1744283392, 31525, 299244432, 22256, 1744285296, 31525])

brng=15 raised an oneMKL error; brng=-1 was accepted as a negative index.
Separately, if (brng): treated brng=0 as unset, but 0 is MT19937:

Intel oneMKL ERROR: Parameter 1 was incorrect on entry to vslGetStreamStateBrng
mkl_random.MKLRandomState(1, brng=0).randint(0, 100, 6)
array([0, 0, 0, 0, 0, 0]) # brng="MT19937" gives [13 56 84 80 76 6]

How found

Coverity Scan CID 653949, OVERRUN: "Overrunning callee's array of size 11 by
passing argument __pyx_v_brng_token (which evaluates to 15) in call to
irk_randomseed_mkl". Confirmed against randomkit.h:50, randomkit.cpp:111 and
the .pyx call path, then reproduced per-process on a build of master. The
brng=0 bug surfaced while writing the boundary test.

Distinct from the _seed_impl OUT_OF_BOUNDS finding already dismissed in
coverity/README.md, which is the tuple unpack and is a false positive.

Fix

• Expose BRNG_KINDS from randomkit.h so the bound tracks the C table.
• Range-check the integer path; out of range warns and falls back to MT19937,
  as an unrecognized generator name already does.
• Use brng is not None instead of truthiness, so brng=0 is honoured.

Tests

5 out-of-range values assert warn plus fallback; 0 and 10 assert accepted,
silent, and equal to the same generator selected by name.

[ndgrigorian]

@vchamarthi Seems it can still be broken for MKLRandomState(seed, brng=None) when seed is None for basically the same reason (irk_get_brng_and_stream_mkl returns -1 and passes to brng_list, so we get OOB read).

We should fix that here too

[vchamarthi]

@vchamarthi Seems it can still be broken for MKLRandomState(seed, brng=None) when seed is None for basically the same reason (irk_get_brng_and_stream_mkl returns -1 and passes to brng_list, so we get OOB read).

We should fix that here too

Fixed in 6d3fa4b. It's broader than seed=None: in __init__ there's no stream yet, so any seed with brng=None hit the -1 index. I made __init__ treat brng=None as "MT19937", the documented default, and added test_init_brng_none_uses_default, which fails before the fix and passes after.

Is defaulting to MT19937 OK, or would you rather brng=None raise a ValueError in the constructor? seed(brng=None) still keeps the current generator.

[ndgrigorian]

Is defaulting to MT19937 OK, or would you rather brng=None raise a ValueError in the constructor? seed(brng=None) still keeps the current generator.

Defaulting to MT19937 seems fine

[ndgrigorian]

LGTM

[antonwolfy]

Missing PR links