nx_orchestrate.nx

nx_orchestrate.nx -- auto lane (eats debt seq163). WAIT-FOR-OPENING QUEUE: sessions QUEUE actions blocked by a sibling's in-flight work instead of colliding. Each knowledge/store/deployq- row = id | title | precondition | status | planid | note precondition = a CHECKABLE predicate, now COMPOUND-capable (v2, 2026-07-20 -- for safe mgmt deploys that must fire only when the fix is STAGED *and* mgmt is CALM): exists:<path> | contains:<path>:<needle> -- single, OR headroom:[<floor_mb>[:<max_centiload>]] -- the box has build headroom NOW (2026-08-18): forks nx_build_admit check [floor] [max]; satisfied iff it GRANTS (exit 0). QUEUE/DENY-MEM/ unreadable-proc all WAIT -- a guard that cannot measure must not fire (fail-closed). This is the wait-for-opening absorber nx_build_admit's own header says its QUEUE verdict was designed for, and that /api/build's REFUSED-LOAD ("nothing is queued for you anywhere") had never been given: a queued build row `headroom: -> plan build-<target>` now fires on the poller pass after the opening appears, instead of a human re-issuing it by hand. <pred> && <pred> -- AND of two (BOTH must hold to fire) (e.g. exists:nx_mgmt_api.elf.new && contains:knowledge/status/api_contract.log:VERDICT=GREEN). This POLLER (cron */5, non-blocking) evaluates every `queued` row: satisfied -> FIRE the row's plan via nx_plan_run (reusing its allowlist+pinned-arg security WHOLESALE -- an action is a pre-seeded plan- workflow, never arbitrary exec) and re-put the row status=fired-rc<N>; unsatisfied -> WAIT (logged, no dead thread, no clobber). SOTA: Temporal durable await-condition, sovereign. Never-brick: fires only plans. nx_orchestrate run [qprefix] [outlog] ENVELOPE v4 (2026-08-29): scan UNCAPPED -- the v3 cap counted SCANNED lines, so 63 dead history rows at the plane head exhausted it and every queued row past index 63 was silently unreachable (measured live: 135 queued rows, fired=0 on every pass, first queued row at plane index exactly 64). A CAP MUST BOUND WORK, NOT SCAN: evaluations of queued rows cap at OC_EVAL_CAP, fires at OC_FIRE_CAP, overflow is COUNTED AND PRINTED (deferred= / fire_deferred=), never dropped in silence. ONE pass at a time (flock; a slow pass makes the next refuse rather than double-fire rows still marked queued). exit: 0 ran (counts printed, or PASS-ALREADY-RUNNING) | 4 queue unreadable | 2 usage. license_tier: ORIGINAL expect_exit: 0