Android prices every fork at ~25ms and every exec at ~90ms, and the
preset apply was paying that price for questions it had already
answered: 95 separate stat calls re-proving file metadata one field at
a time, command substitutions around proc readers whose parsing was
already builtin-only, expected firewall rules rebuilt through captures
on every verify, and printf — an external binary on Android's mksh —
invoked once per line of every publication.
One stat capture now serves every metadata question a proof asks
(path_meta_capture, consumed by the uid/mode/nlink/size predicates and
retired on proof exit); the state-dir proof and a fully-pinned process
identity become lock-scoped facts on the same single-writer argument as
the owner read cache; proc_starttime/proc_argv0 gain fork-free
global-return forms with the printf wrappers kept as seams; the qnum
argument check reuses the cmdline snapshot argv0 already paid for;
expected firewall rules are built as data instead of captured; and
multi-line writers emit once through z2_emit_line, which resolves to
mksh's raw print builtin on the device and printf elsewhere. Load-time
probes replace the per-call command -v PATH walks, and log stamps reuse
one date exec per second via EPOCHREALTIME.
Measured on the Pixel 9 Pro XL: preset apply 56s before this series,
40s after the owner read cache, 14.6s now. tests/shell/run.sh passes;
the three test seams that named replaced internals (proc_starttime,
z2_fw_restore_command, the owner printf format) follow the new forms.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
852d42c02e