nx_dispatch_lease.nx

nx_dispatch_lease.nx -- WMS rung M4: the LEASING DISPATCHER / SCHEDULER. module: nishi-core.autonomy.dispatch_lease capability: DISPATCH_LEASE_NEXT_READY_RUNG (frontier-aware pick + O_EXCL lease) WHAT THIS ADDS (the missing capability): nx_assign_next/an_pick are READ-ONLY -- they NAME the next runnable rung but never CLAIM it, so two parallel workers both see the same "next" and collide on one stream. M4 LEASES the pick under an O_CREAT|O_EXCL lockfile (the kernel guarantees exactly one creator wins -- no TOCTOU), so exactly one owner holds a given stream at a time. Generalizes the conductor's single-arc drill into "keep the WHOLE machine moving": if the top pick is already leased, the dispatcher scans down the w-ordering and leases the NEXT ready stream instead of stalling. REUSE / lineage (compose, do not reinvent): - an_load / an_pick / an_deps_done / an_id_at / an_find <- nx_assign_core (the dep-graph + readiness + frontier-aware pick are ALREADY correct there) - O_EXCL acquire/steal lock <- the rt_lock idiom from nx_registry_lock (INLINED here as dl_lock/dl_unlock so this file keeps a SINGLE nx_syscalls surface and dodges the double-import rc=6 trap -- nx_assign_core already pulls nx_syscalls transitively, importing nx_registry_lock too would surface it twice) - fa_cat / fa_catn / fa_appendz <- nx_framed_append (every status write is ONE buffer -> ONE locked atomic fa_appendz; the torn-line bug we must NOT reintroduce) IMPORT DISCIPLINE: import nx_assign_core (surfaces nx_syscalls) + nx_framed_append (also needs nx_syscalls). The prepass dedups nx_syscalls across imports (proven: nx_assign_next imports BOTH nx_syscalls AND nx_assign_core and builds clean). We do NOT import nx_registry_lock (its O_EXCL logic is inlined) -> only two import edges. license_tier: ORIGINAL