D439 - Cross-section hardware diff: the system-wide divergences worth fixing

measured - 2026-09-01 (user-directed, /loop)

Ran obSCEne's probe under orbistoun and diffed all 262 shared res values against the hardware report. Sorting the ~26 divergences into what transfers and what does not:

System-wide, fixable (a fixed platform behaviour):

• File error codes, deferred by the fs layer until measured and now answered: read/write/lseek on a bad fd = -EBADF (0xffffffff80020009), open of a missing path = ENOENT (0x80020002), open of null = EFAULT (0x8002000e) - orbistoun returns -1 for all (000-boot/write-rejects-bad-fd, 040-file/{lseek,read}-rejects-bad-fd, 040-file/open-rejects-{missing,null}).
• 060-module/dlsym-rejects-bad-handle = 0x80020003 (needs handle validation, D438).
• 005-generation/neo-mode = 1 (a PS5 is neo-class; orbistoun answers 0).
• 165-gnm/dispatch-direct writes 6 dwords, orbistoun writes 5.
• 150-memory-map/{walk,after-allocation} = 0x8/0x9, orbistoun 0x7/0x8 (off by one region).
• The direct-map virtual base 0x2_0000_0000 (020-memory/map, flexible-round-trip), pending its Windows test.

Not transferable - left alone: timing and counters (010-kernel/*, 050-time/usleep, 018-relational/clock), thread and object identities (018-relational/*, 030-thread/self, 100-input/open, 070-user/initial-user - real ids are large where orbistoun uses 1, per-run/account), and the flexible budget (020-memory/flexible-available, per-process - D437). 110-modules/* = 0x20 vs 1 is a real count divergence (hardware sees 32 loaded modules) but reflects a different loaded set, not a fixed value.

Fixing the file error codes first, as the largest coherent cluster.