nx_skullfit_beat.nx

nx_skullfit_beat.nx -- no-arg BEAT: the skull generator's RECURSIVE LEARNING cycle (operator 2026-08-09: "i want nishi ecosystem to gain these functionalities as a recursive learning setup so that it can keep growing and get its capabilities to sota"). Each run advances the whole loop with no seat attached: (1) `nx_skullsdf tune` refines the 76-parameter fit AGAINST THE SCANNED ORACLE, seeded FROM knowledge/skull_fit.dat and saving BACK to it -- successive beats converge instead of restarting; (2) an emit must report fit_src=1 -- the ARTIFACT proves the fit landed, never the tuner's receipt (a builtin-defaults emit means the loop is broken, not neutral); (3) `gapmap` re-measures the per-region gap so the trend line carries WHERE the skull still deviates, not just a scalar; (4) THE CANON GATE measures the FACE, landmark by landmark against the oracle (anatomy AN17: the gapmap's regions are dominated by the vault and were blind to a 49 mm muzzle); (5) ONE line appends to the trend journal. BEAT-FAIL lines are evidence too. ★★★★★★THE WORK PER RUN IS DERIVED FROM THE DISPATCH BOUND, NEVER TYPED (AN17, 2026-09-18). The clock kills any dispatch at knowledge/skullfit_beat.conf dispatch_deadline_ms (the scheduler's own constant, carried as data with its provenance). MEASURED: one tune pass = 477 s on the NAS, so the old three-pass run was killed at 901 s on every dispatch since 2026-08-14 and appended NOTHING -- a beat whose one run cannot fit the deadline is indistinguishable from a beat that never fires. Now every run reads its own previous timings from knowledge/status/skullfit_beat.timing (the MAXIMUM seen per step, so a slow box only makes the plan more conservative), plans passes = floor((bound - emit - gapmap - canon) / pass), at least one, and when even one pass plus the measurements cannot fit it ALTERNATES phases (tune-only, then measure-only) so both still happen. A run stamps `started` before the work and `finished` after: a run the clock killed leaves started without finished, and the NEXT run names it on its trend row (prev_killed=1) and halves its plan -- a named refusal, never silence. The coordinate descent's step ladder (4, 2, 1) is carried ACROSS runs in the same file: a pass that moves the objective keeps its step, a pass that does not halves it. ***VACUOUS-OBJECTIVE GUARD (nx_fitloop's law): final_centi==0 against a real scanned skull is a BROKEN RULER, never perfection -- scored INVALID, exit 2. Scheduler contract: the clockjobs- plane fork+execs this elf with NO args. trend: knowledge/status/skullfit_trend.jsonl line: <epoch>\tTUNE ...\tFIT ...\tGAP ...\tCANON ...\tPLAN ...\n license_tier: ORIGINAL No hw writes (Rule 26). expect_exit: 0