D447 - sysctlbyname answers the knobs it can source, and kern.osrelease stops being refused

measured - 2026-09-01 (user-directed /loop; obSCEne oracle)

135-sysctl/osrelease failed under orbistoun: sysctlbyname had no implementation, so it hit the default stub and refused, and a refused kern.osrelease is precisely what turns firmware detection off in a title that reads it (the sibling-emulator wall obSCEne was built to avoid). Implemented it.

sysctl_value(name, kernel_release) decides what each knob answers, and refuses the rest rather than inventing:

• kern.ostype → "FreeBSD". The one fact this project rests on - the target kernel is FreeBSD-derived (docs/REFERENCES.md) - stated back to the guest, a citable constant, not a guess.
• kern.osrelease → the configured [machine]().kernel_release, NUL-terminated. Empty until a machine sets one, because orbistoun does not invent a kernel version (why Machine::default().kernel_release is empty). A console measured "0.0-prototype" (length 0xe, obSCEne 135-sysctl/osrelease), which a machine profile may carry; the default answers an empty knob rather than pretending to know it.
• everything else → refused. kern.version is a build banner with a measured revision and date, and hw.ncpu a core count obSCEne itself flags the sibling emulator for inventing; orbistoun carries neither honestly yet, so it refuses rather than answers plausibly (principle 3).

The handler follows the POSIX/FreeBSD contract - copies what fits, updates the length, answers the size alone when handed no destination, and never writes past the length it was given (the overrun obSCEne's guard byte catches). A too-small buffer earns ENOMEM, for which errno::NO_MEMORY (12) was added.

Verified against the oracle: 135-sysctl/osrelease now passes (as on the console); 135-sysctl/names improves from refused to partial - it answers kern.ostype and kern.osrelease and honestly refuses the two it cannot source, where hardware answers all four. obSCEne's distinct failures drop from five to four, no new ones. A unit test pins sysctl_value (ostype, a set and an unset release, two refusals). Kernel and core tests pass, clippy clean, new code fmt-clean.

Not the same class, and left: 137-kernelcall/system-version fails under orbistoun but skips on hardware - its gadget needs getpid resolved, which the console refused (0x80020003) so the probe never built it; orbistoun does resolve getpid, reaches the raw syscall and fails it (-78). That is a divergence about resolving and raw syscalls, not a knob to answer, so it waits. Remaining: 110-modules (the one-module gap) and 900-surface/control (the resolver reports a non-existent symbol present).