Parallel-workstream stability — hardware-rung-up, sovereign, measured. Devlog snapshot.
~21 parallel Claude workstreams shared one WSL VM, one /tmp,
and one repo with no isolation and no arbitration. Siblings shredded each other's
/tmp/nxpass mid-handshake, raced the same *.sov.elf builds, and clobbered shared
memory files mid-write. Intermittent push failures and Wsl/Service/E_UNEXPECTED crashes were the
symptom; unsynchronized writes to shared mutable state were the cause.
| Pillar | Sovereign primitive | Gate | Status |
|---|---|---|---|
| Isolation collisions can't happen |
nx_runpath — per-workstream namespace under ~/.nishi/run/<wsid>/, off /tmp |
6/6 | GREEN |
| Arbitration serialize shared sinks |
nx_arbiter — kernel flock lease keyed by named resource |
5/5 | GREEN |
| Stability stop the VM churn |
corrected .wslconfig (evidence-based, from the official MS doc in the no-rot library) |
researched | STAGED |
Not a claim — a reproducible measurement. Four real child processes (our own sys_fork, no shell)
each run 100 read–increment–write cycles on one shared counter file: the exact read-modify-write
that clobbered the memory files. Once unsynchronized, once under the arbiter lease.
| Condition | Final count | Lost updates |
|---|---|---|
| Unlocked (F-class) | 20 / 400 | 380 lost (95%) |
| Arbitrated (S-class) | 400 / 400 | 0 lost |
Self-contained and sovereign: our fork/wait4 + our flock + our counter file. No third party, no shell-driven race.
Gate: nx_arbiter_race_gate 3/3 GREEN.
nx_synced_push — the arbiter-wrapped pusher —
so its delivery is itself the first live exercise of the nas_push lease.Nishi — sovereign from the hardware rung up. nx_cc→nxasm, no gcc. Read-only devlog snapshot.