PR_SET_DUMPABLE now stores 0 or 1 and PR_GET_DUMPABLE returns that state. Fork inherits it. An ordinary exec resets it to 1; mixed real/effective user or group credentials select 0. Effective user/group changes also select 0. A nondumpable process's maps, smaps and auxv are protected from other processes with matching credentials; the process itself and CAP_SYS_PTRACE retain access.

Build the desired kernel, then run tests/dumpability/run or tests/dumpability/run --arch amd64. The runner accepts the same isolated-worktree environment variables as tests/network-options. The guest checks retained values, invalid inputs, fork/exec, credential resets and actual procfs reads from a same-UID peer.

Mixed real/effective user or group credentials also emit AT_SECURE=1 in the initial auxiliary vector, with matching UID/EUID/GID/EGID entries. Ordinary matching credentials emit zero. The guest checks these after real execs on both architectures. This closes the existing-credential secure-loader part of SEC4; executable set-ID and file-capability transitions remain separate work.

The independent guest is maintained in dumpfixture/core.v, using native libc declarations from its ABI header. Both real exec child modes and all 14 original check sites are retained, producing 15 verdicts across the UID/GID cases. The runner compiles the selected V target and preserves its 300-second budget.

This addresses the dumpability portion of OB2. ELF core writing, RLIMIT_CORE output limits, privileged dump destinations and file-capability/set-ID exec integration remain to be implemented; this change does not create core files.

The native V controller accepts --state-dir FRESH-DIR to retain artifacts and --build-only to qualify the fixture separately. Every run snapshots an already-built matching kernel.