code wiki / _hdl_build / nx_clock_sched.nx

nx_clock_sched.nx source

↩ module page · 1207 lines · 77255 B

1// nx_clock_sched.nx -- RENAMED FROM nx_clock.nx 2026-07-31 (lib-reconcile). PERMANENT FIX for a 2// two-libraries-one-name collision: runtime/nx_clock.nx is the TIMING library (nx_clock_monotonic_ns, 3// 62 importers); THIS file is an unrelated tickless JOB SCHEDULER (clk_* registry + dispatcher) that 4// only shared the filename. nx_cc binds an import to the IMPORTER'S OWN DIRECTORY FIRST, so every 5// _hdl_build organ importing nx_clock.nx silently got the SCHEDULER -- _clk_probe.nx, whose whole 6// purpose is to prove nx_clock_monotonic_ns works, could not resolve it. The RENAME is the fix. 7// so we avoid issues ... dont have a million pulses and daemons"). Researched (knowledge/fetched/sched_*.raw: 8// clock-distribution = ONE oscillator -> a tree of DIVIDERS; PLL derives every frequency from ONE reference; 9// cron = ONE daemon reads ONE table of timed jobs; tickless = don't burn a constant tick when idle). The S-class 10// pattern is identical: ONE tick source + ONE durable JOB REGISTRY + ONE dispatcher. Every periodic capability 11// REGISTERS a job (name, interval-in-ticks) = a divider off the single clock -- it does NOT spin up its own 12// daemon/pulse loop. So N capabilities cost ONE loop, not N. This library is the registry + dispatcher; the single 13// tick source drives it (one external spark, like a crystal). license_tier: ORIGINAL 14import "nx_syscalls.nx" 15import "nx_store_seed_lib.nx" 16import "nx_clock_caps.nx" 17import "nx_ioadmit_lib.nx" // ioa_measure / ioa_spawn_budget: THE I/O-storm ruler nx_build_admit uses -- composed here so the clock and the build gate can never disagree about whether the box is in a storm (2026-09-02) 18const CLK_MAGIC_1000000: i64 = 1000000 19 20const CLK_MAXJOBS: i64 = 512 // WAS 128 (2026-08-19): the desired plane hit 128/128 and every 21 // over-cap path was a SILENT drop -- a beat that vanishes at merge. 22 // 4x headroom; clk_register now refuses LOUDLY at the cap. Residual, 23 // named: the inline merge-at-cap skips stay quiet above 512. 24// DUTY-CYCLE CEILING FOR A COST-OVERRUNNING JOB (2026-08-06). Re-arming an overrunning job to exactly 25// its own runtime still leaves it consuming HALF the serial scheduler -- MEASURED: nx_worldgen_gate runs 26// 289s against a declared 120s period, so plain backoff took the blackout from ~96pct to ~50pct and 27// sitecheck stayed frozen through every run. Backing off to runtime*9 caps any single job at ~1/10 of 28// the scheduler. NOT A MAGIC NUMBER: it is the reciprocal of the duty-cycle budget (9 => <=10pct), and 29// it is the ONLY knob here, so raising it loosens the ceiling monotonically and nothing else moves. 30// (STAR)A SERIAL SCHEDULER MUST BOUND WHAT FRACTION OF ITSELF ONE JOB CAN OWN, BECAUSE EVERY OTHER JOB'S 31// LIVENESS IS THAT FRACTION'S COMPLEMENT. 32const CLK_HOGDUTY: i64 = 9 33// ---- PER-DISPATCH DEADLINE (2026-08-06): ADOPTION, NOT INVENTION ------------------------------- 34// nx_guarded_run.nx ALREADY solved "fork + wait4 WNOHANG + deadline + SIGKILL" and a dozen organs use 35// it (nx_tool_run, nx_sweep_core, nx_boot_revive, nx_web_crawl_step...). The clock simply never adopted 36// it and kept a bare BLOCKING sys_wait4(pid, st, 0) -- which is why one child could own the scheduler 37// indefinitely (debt 1786056459; proven live: 475s+ stall and the 1800s budget overrun by 42s). 38// (STAR)THE ESTATE HAD ALREADY SOLVED IT AND THE CLOCK DID NOT LOOK -- an ADOPTION gap, not a missing 39// primitive. Inlined rather than imported because nx_guarded_run.nx carries its own main() self-test. 40// nx_kill is the SANCTIONED wrapper: nx_syscalls notes the compiler BAKES bodies BY NAME, so a function 41// literally named sys_kill emits syscall 8; nx_kill passes rv64 129 which the sovereign table maps to 42// x86_64 62. Using the raw x86 62 here would have silently become lseek(8) and never killed anything. 43// DEADLINE = ONE FULL WINDOW: a single dispatch longer than the entire window is starving that window 44// BY DEFINITION, and it reaps nothing measured legitimate (worldgen 289s/286s, segguard ~700s). 45const CLK_SIGKILL: i64 = 9 46const CLK_DISPATCH_POLL_MS: i64 = 250 47// PER-JOB KILL DEADLINE. 900000ms = 900s = HALF the 1800s scheduling window (TLN_WINDOW_SECS). 48// WAS 1800000ms = the ENTIRE window, which is not a deadline at all: a single job could legally consume 49// the whole scheduler and the guard could not fire before the window ended anyway. A BOUND EQUAL TO THE 50// BUDGET CANNOT BIND. The call site below sells this as the fix for "the single line that let one child 51// own the clock" -- the bounded wait DID replace a bare blocking wait4, but the bound was then set to the 52// full budget, so the protection was nominal. 53// DERIVED, NOT CHOSEN: half the window means no single job can ever take more than half the scheduler, 54// and it carries 2.2x margin over the largest LEGITIMATE runtime measured on 2026-08-14 (max_exec across 55// three consecutive windows: 41s, 92s, and 408s for nx_compare_beat.elf). If that margin is ever wrong 56// the failure is VISIBLE, not silent: a job killed here is logged by clk_actlog with code 124 (the shell 57// timeout convention), so an over-tight bound surfaces as a named row rather than a mystery. 58// HONEST SCOPE: this bounds the PATHOLOGICAL case only. It does NOT fix the starvation measured that day 59// (b=11 -> 16 -> 23), which is cumulative time on a SERIAL dispatcher -- exec 761s -> 990s -> 1167s of an 60// 1800s window -- not one job monopolising. That needs bounded concurrency in clk_run_edf (debt 3952). 61const CLK_DISPATCH_DEADLINE_MS: i64 = 900000 62const CLK_NAMEW: i64 = CLK_CMD_CAP // bytes per name/organ slot -- DERIVED from the ONE shared 63 // command cap in nx_clock_caps.nx, the SAME const nx_clockjob's 64 // CJ_ROWCAP is derived from, so the reader can never hold less than 65 // the writer admits. WAS 128 (2026-08-03..08-22): every command 66 // longer than 127 bytes was SILENTLY CUT at merge/save/load -- 7 67 // live rows executed truncated argv (gateroster beat ran with 68 // deadline 1800 instead of 180000; admit/refusalshape/alertscore 69 // wrote knowledge/gateroster.c, knowledge/stat and kno). Before 70 // that it WAS 48 (debt 1784413227): 51/63-char paths fork-failed 71 // 127 on every tick. A SLOT NARROWER THAN WHAT THE WRITER ADMITS 72 // IS A SILENT TRUNCATOR WEARING A CONSTANT -- bind both ends to 73 // ONE imported const, never mirror a number. 74const CLK_ROWW: i64 = 2*CLK_NAMEW + 64 // per-row serialization budget, DERIVED, not chosen: 75 // 2 slots + 2 i64 decimals (<=20 bytes each) + 3 tabs + newline 76 // = 2*CLK_NAMEW + 44, held with 64 so the two numbers can never 77 // outgrow it (a hand-counted width beside a widened slot is the 78 // drift class this whole change retires) 79const CLK_REG: *u8 = "knowledge/sched/jobs.tsv" // RETIRED LEGACY PATH (2026-08-03, debt 1785792856): 80 // kept only so old fixtures parse; the live state SSOT is the 81 // clocksched- seg-store plane (clk_load_state / clk_save_plane below) 82 83// ---- STORM-AWARE DISPATCH FOR HEAVY BEATS (2026-09-02, /compare/loadgov LV14) ---------------------- 84// MEASURED THE SAME DAY: procchurn read cpu busy 24.5 percent with blocked >> running (I/O-bound), the 85// D-state roster carried nx_web_crawl_step (706 s), nx_secret_scan and nx_web_shard_compact -- all CLOCK 86// BEATS -- and every seat build was refused by admission for the same storm those beats were feeding. 87// This dispatcher had NO admission at all: a due job forked regardless of the box, so the beats that 88// cause a storm fired INTO it while the one lane that does check admission (the build queue) waited. 89// A SCHEDULER THAT CHECKS NOTHING IS THE STORM'S FEEDER, AND THE POLITE LANE PAYS FOR IT. 90// DESIGN, data-driven and fail-open: a job is HEAVY only if its NAME is listed in CLK_HEAVY_CONF (one 91// exact name per line; an unlisted job dispatches byte-for-byte as before, and an absent conf lists 92// nothing). A heavy job that is due while the shared ruler reports ZERO spawn budget (procs_blocked at 93// the storm line, the same conjunct nx_build_admit refuses on) is DEFERRED: re-armed to now plus a 94// fraction of its own period, counted, and announced on the in-flight beat as deferred-storm. It still 95// runs -- one storm cannot park it past CLK_DEFER_MAX_S per probe -- and EDF's own starvation report 96// still names it if the deferral compounds. An UNREADABLE ruler never defers: an axis that cannot see 97// must abstain, and here abstaining means the pre-change behaviour, never a stalled clock. 98// The instruments themselves (resmon, memvel, procchurn, sitecheck, netobs, tlsprobe, kaprobe) must 99// NEVER be listed: a detector deferred by the storm it detects is the axis-blind defect wearing a conf. 100const CLK_HEAVY_CONF: *u8 = "knowledge/status/clock_heavy.conf" 101const CLK_DEFER_DIV: i64 = 4 // re-arm at interval/4: a 1800 s beat re-checks every 450 s, so a storm costs it at most 3 probes before its own next period 102const CLK_DEFER_MIN_S: i64 = 60 // floor: the shortest re-check that is not a busy loop against /proc/stat 103const CLK_DEFER_MAX_S: i64 = 900 // cap: half a window, the same bound CLK_DISPATCH_DEADLINE_MS already places on one job 104// PURE: exact-whole-line membership of `name` in a newline-separated conf (CR-tolerant), the same 105// grammar the tools daemon's async_only_tools.conf reader uses, so a name with a suffix can never match. 106func clk_heavy_listed(conf: *u8, n: i64, name: *u8) -> i64 { 107 if n <= 0 { return 0 } 108 var nl: i64 = 0 109 while name[nl] != (0 as u8) { nl = nl + 1 } 110 if nl == 0 { return 0 } 111 var i: i64 = 0 112 var hit: i64 = 0 113 while i < n { 114 var e: i64 = i 115 var eol: i64 = 0 116 while eol == 0 { 117 if e >= n { eol = 1 } else { if conf[e] == (10 as u8) { eol = 1 } else { e = e + 1 } } 118 } 119 var len: i64 = e - i 120 if len > 0 { if conf[i + len - 1] == (13 as u8) { len = len - 1 } } 121 if len == nl { if hit == 0 { 122 var k: i64 = 0 123 var same: i64 = 1 124 while k < nl { if conf[i + k] != name[k] { same = 0; k = nl } else { k = k + 1 } } 125 if same == 1 { hit = 1 } 126 } } 127 i = e + 1 128 } 129 return hit 130} 131// PURE: how long a deferred heavy job waits before EDF may pick it again -- a fraction of its own 132// period, clamped so a 60 s beat is not busy-polled and a daily beat is not parked for hours. 133func clk_defer_secs(interval: i64) -> i64 { 134 var d: i64 = interval / CLK_DEFER_DIV 135 if d < CLK_DEFER_MIN_S { d = CLK_DEFER_MIN_S } 136 if d > CLK_DEFER_MAX_S { d = CLK_DEFER_MAX_S } 137 return d 138} 139// PURE: the decision. heavy=1 AND budget==0 -> DEFER. budget<0 is UNOBSERVABLE (the ruler could not 140// read /proc/stat) and must NOT defer: abstain toward the pre-change behaviour. A light job never defers. 141func clk_storm_defer(heavy: i64, budget: i64) -> i64 { 142 if heavy != 1 { return 0 } 143 if budget == 0 { return 1 } 144 return 0 145} 146// The live measurement, composed from the shared ruler with its OWN calibration constants, so the 147// clock refuses on exactly the D-state line the build gate refuses on. ioa[0]=ncpu ioa[1]=procs_blocked. 148func clk_storm_budget(ioa: *i64) -> i64 { 149 if ioa_measure(ioa) != 0 { return IOA_UNREADABLE } 150 return ioa_spawn_budget(ioa[0], ioa[1], IOA_BLOCKED_PER_CPU, IOA_RESERVE_SLOTS) 151} 152 153func clk_slot(names: *u8, i: i64) -> *u8 { return ((names as i64) + i*CLK_NAMEW) as *u8 } 154func clk_streq(a: *u8, b: *u8) -> i64 { var i: i64=0; while a[i]!=(0 as u8){ if a[i]!=b[i]{return 0} i=i+1 } if b[i]!=(0 as u8){return 0} return 1 } 155 156func clk_find(names: *u8, n: i64, name: *u8) -> i64 { var i: i64=0; while i<n { if clk_streq(clk_slot(names,i), name)==1 { return i } i=i+1 } return 0-1 } 157 158// REGISTER a periodic job = add a divider off the one clock. Idempotent: re-registering the same name is a no-op 159// (so capabilities can declare their job every boot without ever creating a duplicate pulse). returns 1 if added. 160func clk_register(names: *u8, organs: *u8, intervals: *i64, next_due: *i64, np: *i64, name: *u8, organ: *u8, interval: i64) -> i64 { 161 let n: i64 = np[0] 162 // A re-registration whose ORGAN or INTERVAL differs must LAND. This returned 0 for both 163 // "already present, identical" and "already present, DIFFERENT", so every correction was a 164 // silent no-op. Measured over the whole clockjobs- plane 2026-08-07: four names carry different 165 // commands across rows and the live beat was the OLD one in every case -- segguard still ran 166 // nx_segguard.sh while nx_segsweep.elf had been re-registered three times; lmexport, stalesweep 167 // and ddqbeat likewise still ran their .cron.sh. A shell-to-organ migration of four beats was 168 // written and never took effect, and nothing could see it because 0 was the only signal. 169 // An IDENTICAL re-declaration stays a no-op, so capabilities may still declare their job every 170 // boot without churn. Returns 0 unchanged, 2 updated. 171 // ★★★★★★ IDEMPOTENCE AND IMMUTABILITY ARE NOT THE SAME PROPERTY: AN IDEMPOTENT-BY-NAME 172 // REGISTRAR MAKES THE FIRST WRITE PERMANENT AND EVERY LATER CORRECTION A SILENT NO-OP. 173 let ex: i64 = clk_find(names, n, name) 174 if ex >= 0 { 175 var iv0: i64 = interval 176 if iv0 < 1 { iv0 = 1 } 177 var eo: *u8 = clk_slot(organs, ex) 178 var same: i64 = 1 179 var q: i64 = 0 180 var st: i64 = 0 181 while st == 0 { 182 let a: i64 = eo[q] as i64 183 let b: i64 = organ[q] as i64 184 if a != b { same = 0; st = 1 } 185 if st == 0 { 186 if a == 0 { st = 1 } 187 if a != 0 { 188 q = q + 1 189 if q >= CLK_NAMEW { st = 1 } 190 } 191 } 192 } 193 if intervals[ex] != iv0 { same = 0 } 194 if same == 1 { return 0 } 195 var w: i64 = 0 196 while w < CLK_NAMEW - 1 { 197 if organ[w] == (0 as u8) { eo[w] = 0 as u8; w = CLK_NAMEW } else { eo[w] = organ[w]; w = w + 1 } 198 } 199 if w == CLK_NAMEW - 1 { eo[w] = 0 as u8 } 200 intervals[ex] = iv0 201 next_due[ex] = iv0 202 return 2 203 } 204 if n >= CLK_MAXJOBS { 205 // A CAP REACHED IN SILENCE BECOMES A REGISTRATION NOBODY KNOWS WAS DROPPED (2026-08-19: 206 // measured 128/128 with every over-cap path a quiet return). Say WHICH row was refused. 207 let cm: *u8 = "CLOCK-CAP-DROP register refused at CLK_MAXJOBS: " as *u8 208 var cl: i64 = 0 209 while cm[cl] != (0 as u8) { cl = cl + 1 } 210 sys_write(1, cm, cl) 211 var nl2: i64 = 0 212 while name[nl2] != (0 as u8) { nl2 = nl2 + 1 } 213 sys_write(1, name, nl2) 214 sys_write(1, "\n" as *u8, 1) 215 return 0 216 } 217 var d: *u8 = clk_slot(names, n); var i: i64=0 218 while i<CLK_NAMEW-1 { if name[i]==(0 as u8){ d[i]=0 as u8; i=CLK_NAMEW } else { d[i]=name[i]; i=i+1 } } 219 if i==CLK_NAMEW-1 { d[i]=0 as u8 } 220 var e: *u8 = clk_slot(organs, n); var j: i64=0 221 while j<CLK_NAMEW-1 { if organ[j]==(0 as u8){ e[j]=0 as u8; j=CLK_NAMEW } else { e[j]=organ[j]; j=j+1 } } 222 if j==CLK_NAMEW-1 { e[j]=0 as u8 } 223 var iv: i64 = interval; if iv < 1 { iv = 1 } 224 intervals[n] = iv; next_due[n] = iv; np[0] = n + 1 225 return 1 226} 227 228// THE DISPATCHER: advance to logical tick `now`; mark + return how many jobs are DUE (next_due<=now), advancing 229// each due job's next_due by its interval (catch-up safe: a job never fires more than once for a missed window). 230func clk_tick(intervals: *i64, next_due: *i64, n: i64, now: i64, fired: *i64) -> i64 { 231 var count: i64 = 0; var i: i64 = 0 232 while i < n { 233 if next_due[i] <= now { 234 fired[i] = 1; count = count + 1 235 while next_due[i] <= now { next_due[i] = next_due[i] + intervals[i] } // re-arm past now (no runaway on a skipped tick) 236 } else { fired[i] = 0 } 237 i = i + 1 238 } 239 return count 240} 241 242// THE FUNCTIONAL DISPATCH: advance to `now` and actually RUN each due job by fork+exec of its organ elf path 243// (the same fork+exec idiom nx_god_pulse / nx_aw_hostctl use). One dispatcher runs N jobs at their divided rates; 244// there is NO per-job daemon. Parent waits each child so a slow job can't be lost (a real scheduler can make this 245// bounded-concurrent / fire-and-forget). `organs` is a parallel slot array (organ[i] = an executable path). Returns 246// the number of jobs dispatched this tick. 247func clk_dispatch_run(organs: *u8, intervals: *i64, next_due: *i64, n: i64, now: i64, fired: *i64) -> i64 { 248 var count: i64 = 0; var i: i64 = 0 249 while i < n { 250 if next_due[i] <= now { 251 fired[i] = 1 252 while next_due[i] <= now { next_due[i] = next_due[i] + intervals[i] } // re-arm past now (catch-up safe) 253 let path: *u8 = clk_slot(organs, i) 254 __syscall(90, path as i64, 0x1ed, 0, 0, 0, 0) // chmod 0755 first: recv-shipped organs aren't reliably +x (HC_MGMT_CMD chmods for the same reason) -> without this execve fails 127 = silent no-dispatch 255 let pid: i64 = sys_fork() 256 if pid == 0 { 257 let argv: *i64 = sys_mmap(32) as *i64; argv[0] = path as i64; argv[1] = 0 258 let envp: *i64 = sys_mmap(16) as *i64; envp[0] = "PATH=/usr/bin:/bin" as *u8 as i64; envp[1] = 0 259 sys_execve(path, argv, envp) 260 sys_exit(127) 261 } 262 let st: *i64 = sys_mmap(16) as *i64; st[0] = 0 263 sys_wait4(pid, st, 0) 264 let raw: i64 = st[0]; let sig: i64 = raw & 0x7f; let code: i64 = (raw >> 8) & 0xff 265 // HONEST count: a job whose organ failed to exec (child exit 127 = missing/broken organ) did NOT run, 266 // so it is not counted as dispatched (it gets fired[]=1 so a caller can flag it, mirroring the 267 // publisher's dead-letter). A crash (signalled) or any other exit means the organ DID run. 268 // NO JOB NAME ON THIS PATH, BY CONSTRUCTION (2026-09-18): clk_dispatch_run holds no names array and 269 // its signature is shared with nx_clock_run and nx_clock_dispatch_gate, so it passes null and its 270 // row stays byte-identical to the old shape. The LIVE path (clk_run_edf -> clk_dispatch_one) names the job. 271 clk_actlog(path, code, 0 as *u8) 272 if sig != 0 { count = count + 1 } else { if code != 127 { count = count + 1 } } 273 } else { fired[i] = 0 } 274 i = i + 1 275 } 276 return count 277} 278 279// TICKLESS run (sched_tickless lesson): from start_tick, run up to maxbeats beats, but before each beat SLEEP 280// exactly until the MINIMUM next_due across all jobs -- skipping every idle tick -- then dispatch the due organs. 281// Advances next_due[] in place (the caller persists for resume). out[0..3] = beats, dispatches, skipped_idle, 282// final_tick. ONE implementation, shared by the driver (nx_clock_tickless) and its gate (no parallel copy). 283func clk_run_tickless(organs: *u8, intervals: *i64, next_due: *i64, n: i64, start_tick: i64, maxbeats: i64, tick_ms: i64, out: *i64) -> i64 { 284 let fired: *i64 = sys_mmap(CLK_MAXJOBS*8) as *i64 285 var T: i64 = start_tick; var beats: i64 = 0; var disp: i64 = 0; var skipped: i64 = 0 286 while beats < maxbeats { 287 var minnd: i64 = next_due[0]; var i: i64 = 1 288 while i < n { if next_due[i] < minnd { minnd = next_due[i] } i = i + 1 } 289 if minnd <= T { minnd = T + 1 } 290 let skip: i64 = minnd - T 291 if skip > 1 { skipped = skipped + (skip - 1) } 292 let totms: i64 = skip * tick_ms 293 let ts: *i64 = sys_mmap(16) as *i64; ts[0] = totms / 1000; ts[1] = (totms - (totms/1000)*1000) * CLK_MAGIC_1000000 294 __syscall(35, ts as i64, 0, 0, 0, 0, 0) // nanosleep until the next due event (tickless) 295 T = minnd 296 disp = disp + clk_dispatch_run(organs, intervals, next_due, n, T, fired) 297 beats = beats + 1 298 } 299 out[0] = beats; out[1] = disp; out[2] = skipped; out[3] = T 300 return disp 301} 302 303// STATUS helpers (the consolidation payoff: ONE place shows every periodic job). clk_due_in = ticks until a job 304// fires (<=0 means due now). clk_twin_ok = 1 iff the job's organ elf actually exists (a 0 means it is registered 305// but un-blessed -> the dispatcher would honestly report it didn't run -> surface it in status, do not hide it). 306func clk_due_in(next_due_i: i64, now: i64) -> i64 { return next_due_i - now } 307func clk_twin_ok(organ: *u8) -> i64 { let fd: i64 = sys_openat_rd(organ); if fd < 0 { return 0 } sys_close(fd); return 1 } 308 309func clk_itoa(buf: *u8, o: i64, v: i64) -> i64 { var w: i64=o; var m: i64=v; if m==0{buf[w]=48 as u8;return w+1} if m<0{buf[w]=45 as u8;w=w+1;m=0-m} let t:*u8=sys_mmap(24); var k:i64=0; while m>0{t[k]=(48+(m%10)) as u8;m=m/10;k=k+1} var j:i64=0; while j<k{buf[w]=t[k-1-j];w=w+1;j=j+1} return w } 310 311// persist the registry (the ONE crontab): name<TAB>interval<TAB>next_due per line. atomic via tmp+rename. 312func clk_msgcat(d: *u8, o: i64, s: *u8) -> i64 { var x: i64=o; var i: i64=0; while s[i]!=(0 as u8){d[x]=s[i];x=x+1;i=i+1} return x } 313func clk_msgnum(d: *u8, o: i64, v: i64) -> i64 { 314 var x: i64 = o 315 var m: i64 = v 316 if m < 0 { d[x] = 45 as u8; x = x + 1; m = 0 - m } 317 if m == 0 { d[x] = 48 as u8; return x + 1 } 318 let t: *u8 = sys_mmap(24) 319 var k: i64 = 0 320 while m > 0 { t[k] = (48 + (m % 10)) as u8; m = m / 10; k = k + 1 } 321 var j: i64 = 0 322 while j < k { d[x] = t[k - 1 - j]; x = x + 1; j = j + 1 } 323 sys_munmap(t, 24) 324 return x 325} 326// ADOPTION VISIBILITY (2026-08-07). The clock captured per-job stdout to logs/<job>.log but wrote 327// NOTHING to knowledge/status/actlog.jrnl -- the ledger nx_catalog reads to decide INVOKED. 328// MEASURED: 19 REGISTERED TOOLS ARE CLOCK-DRIVEN, so every one of them could ONLY EVER report 329// REGISTERED-DARK, callable and authorised and NEVER RUN, no matter how often it actually ran. The 330// estate adoption headline was structurally blind to its entire scheduled population -- not wrong 331// about a few organs, blind to a whole class. 332// The row carries the ORGAN command, NOT the job name, because nx_catalog matches the ORGAN via a 333// substring scan of the ledger and the two names differ (job raidwatch vs organ nx_raidwatch). 334// FAIL-OPEN BY CONSTRUCTION: if the ledger cannot be opened this returns immediately and the beat is 335// untouched. Telemetry must never be able to stop the clock -- the same rule the per-job log capture 336// already follows. Tabs and newline are written as BYTES so no source escape can be mangled. 337// 338// JOB NAME ON THE ROW (2026-09-18). The row named only the ORGAN, so a clock row that runs 339// `nx_plan_run.elf <plan>` logged nx_plan_run.elf and nothing else, and the close ruler (nx_rungclose) -- 340// which judges a rung OPERATED when its agent, a clock ROW NAME such as twincheck or searchship-dry, has 341// run since the rung landed -- could never witness ANY plan-run agent. The row now appends ` job=<row 342// name>` as the LAST token of its LAST field: `... clockjob lane=clock exit=<code> job=<name>`. ADDITIVE: 343// still six TAB fields, the ORGAN column byte-identical, every pre-existing byte in place. 344// THE LINE BUDGET grows by exactly one NAME slot, DERIVED, never chosen: the ORGAN (<= one slot) + the 345// JOB name (<= one slot) + <= 88 fixed bytes (two i64 decimals at <= 20 each, the literal words, five 346// TABs, " job=" and the LF), held with 128 so the fixed part can never outgrow it. 347const CLK_ACTLOGW: i64 = CLK_NAMEW + CLK_NAMEW + 128 348const CLK_TOKEN_MIN_BYTE: i64 = 33 // the first printable ASCII byte that is not a space; every byte below it is a space or a control byte 349const CLK_TOKEN_DEL_BYTE: i64 = 127 // ASCII DEL, the one control byte above the printable range 350 351// A JOB NAME IS WRITTEN ONLY WHEN IT CAN TRAVEL AS ONE TOKEN. Returns its length (1..CLK_NAMEW-1), or 0 352// for a null, empty, slot-overflowing or non-token name -- and on 0 the row carries NO job= field, the 353// pre-2026-09-18 shape byte for byte. WHY ABSENCE AND NOT A PLACEHOLDER: the reader matches `job=<agent>` 354// token-bounded, so an empty value, a value a SPACE splits in two, or a TAB/CR/LF that breaks the row 355// grammar could each hand it a witness for a job that never ran; a missing field can only read as "not 356// witnessed", the direction that under-claims. The scan is BOUNDED BY THE SLOT, so a name with no NUL 357// inside its slot can never run the line past CLK_ACTLOGW. Bytes >= 128 pass (a UTF-8 name is a token). 358func clk_job_token_len(job: *u8) -> i64 { 359 if (job as i64) == 0 { return 0 } 360 var n: i64 = 0 361 var st: i64 = 0 362 while st == 0 { 363 if n >= CLK_NAMEW - 1 { st = 1 } 364 else { 365 let c: i64 = job[n] as i64 366 if c == 0 { st = 1 } 367 else { 368 if c < CLK_TOKEN_MIN_BYTE { return 0 } 369 if c == CLK_TOKEN_DEL_BYTE { return 0 } 370 n = n + 1 371 } 372 } 373 } 374 if job[n] != (0 as u8) { return 0 } 375 return n 376} 377 378// THE LEDGER ROW, AS A PURE FUNCTION (2026-09-18). Composes one actlog.jrnl row into `ln` (>= CLK_ACTLOGW 379// bytes) and returns its length. Split out of clk_actlog for the reason clk_split_argv was: a gate can 380// prove the row's shape WITHOUT writing the production ledger, and it proves it on the SAME code the 381// dispatcher runs -- one implementation, no mirrored copy to drift. 382// SHAPE: <epoch> TAB clock TAB <organ> TAB run TAB ok|fail TAB clockjob lane=clock exit=<code>[ job=<name>] LF 383func clk_actlog_row(ln: *u8, now: i64, organ: *u8, code: i64, job: *u8) -> i64 { 384 var o: i64 = clk_msgnum(ln, 0, now) 385 ln[o] = 9 as u8; o = o + 1 386 o = clk_msgcat(ln, o, "clock" as *u8) 387 ln[o] = 9 as u8; o = o + 1 388 o = clk_msgcat(ln, o, organ) 389 ln[o] = 9 as u8; o = o + 1 390 o = clk_msgcat(ln, o, "run" as *u8) 391 ln[o] = 9 as u8; o = o + 1 392 if code == 0 { o = clk_msgcat(ln, o, "ok" as *u8) } else { o = clk_msgcat(ln, o, "fail" as *u8) } 393 ln[o] = 9 as u8; o = o + 1 394 o = clk_msgcat(ln, o, "clockjob lane=clock exit=" as *u8) 395 o = clk_msgnum(ln, o, code) 396 if clk_job_token_len(job) > 0 { 397 o = clk_msgcat(ln, o, " job=" as *u8) 398 o = clk_msgcat(ln, o, job) 399 } 400 ln[o] = 10 as u8; o = o + 1 401 return o 402} 403 404// `job` is the clock ROW NAME (clk_slot(names, i)); null or empty writes the row without a job= field. 405func clk_actlog(organ: *u8, code: i64, job: *u8) -> i64 { 406 var fd: i64 = sys_openat_append("knowledge/status/actlog.jrnl" as *u8, 420) 407 if fd < 0 { fd = sys_openat_append("/volume1/homes/elderwesto/nishihost/knowledge/status/actlog.jrnl" as *u8, 420) } 408 if fd < 0 { return 0 } 409 let ln: *u8 = sys_mmap(CLK_ACTLOGW) 410 let o: i64 = clk_actlog_row(ln, sys_now_realtime_sec(), organ, code, job) 411 sys_write(fd, ln, o) 412 sys_close(fd) 413 sys_munmap(ln, CLK_ACTLOGW) 414 return 0 415} 416func clk_save(path: *u8, names: *u8, organs: *u8, intervals: *i64, next_due: *i64, n: i64) -> i64 { 417 // ---- RELOAD-MERGE BEFORE SAVE (2026-07-31, id 1785559048) ------------------------------------------- 418 // PROVEN CLOBBER, by controlled experiment not inference: a well-formed row appended to the registry via 419 // the sovereign write path returned OK and read back present IMMEDIATELY -- and was GONE 75 seconds later, 420 // with no error, no log and no rejection. Reproduced twice. Cause is right here: this function serialized 421 // ONLY the caller's in-memory set, so the registry could contain nothing but what THIS process happened to 422 // load, and every row added between our load and our save was erased by the atomic rename below. 423 // WHY IT MATTERED: it silently closed what was then the only working extension point for scheduling. 424 // (⚠A CLAIM HERE WAS STALE AND IS RETRACTED 2026-08-03: this comment used to assert the clockjobs- 425 // plane was "retired/dead". MEASURED FALSE -- the live tickless clock merges that plane every window 426 // and rows registered via `nx_store_put knowledge/store/clockjobs- put <actor> <name> <interval> 427 // <organ>` demonstrably arm and fire. The plane IS the sanctioned add lane.) 428 // THE REMEDY IS THE ONE sts_append_row ALREADY APPLIED TO THE SEG-STORE: re-read the file we are about to 429 // overwrite and keep whatever we did not know about. Rows WE hold win outright (ours carry the advanced 430 // next_due); a row we have never seen is appended verbatim, keeping its own next_due so it arms exactly 431 // when its author intended instead of being silently re-armed or dropped. 432 // u00e2u02dcu2026A WRITER THAT SERIALIZES ONLY ITS OWN MEMORY CANNOT COEXIST WITH ANY OTHER WRITER. 433 // u00e2u02dcu2026AN EXTENSION POINT THAT SILENTLY DISCARDS EXTENSIONS IS A CLOSED SYSTEM WEARING OPEN DOCUMENTATION. 434 var nn: i64 = n 435 let rb: *u8 = sys_mmap(CLK_MAXJOBS*CLK_ROWW) 436 let nm2: *u8 = sys_mmap(CLK_NAMEW) 437 let og2: *u8 = sys_mmap(CLK_NAMEW) 438 let rfd: i64 = sys_openat_rd(path) 439 if rfd >= 0 { 440 var rn: i64 = 0 441 var rr: i64 = 1 442 while rr > 0 { 443 rr = sys_read(rfd, ((rb as i64) + rn) as *u8, CLK_MAXJOBS*CLK_ROWW - rn) 444 if rr > 0 { rn = rn + rr } 445 } 446 sys_close(rfd) 447 var p: i64 = 0 448 var ls: i64 = 0 449 while p <= rn { 450 var eol: i64 = 0 451 if p == rn { eol = 1 } else { if rb[p] == (10 as u8) { eol = 1 } } 452 if eol == 1 { 453 if p > ls { 454 var f: i64 = 0 455 var q: i64 = ls 456 var iv2: i64 = 0 457 var nd2: i64 = 0 458 var w2: i64 = 0 459 var g2: i64 = 0 460 while q < p { 461 if rb[q] == (9 as u8) { f = f + 1 } else { 462 if f == 0 { if w2 < CLK_NAMEW - 1 { nm2[w2] = rb[q]; w2 = w2 + 1 } } 463 else { if f == 1 { if rb[q] >= (48 as u8) { if rb[q] <= (57 as u8) { iv2 = iv2*10 + ((rb[q] - (48 as u8)) as i64) } } } 464 else { if f == 2 { if rb[q] >= (48 as u8) { if rb[q] <= (57 as u8) { nd2 = nd2*10 + ((rb[q] - (48 as u8)) as i64) } } } 465 else { if g2 < CLK_NAMEW - 1 { og2[g2] = rb[q]; g2 = g2 + 1 } } } } 466 } 467 q = q + 1 468 } 469 nm2[w2] = 0 as u8 470 og2[g2] = 0 as u8 471 if w2 > 0 { if g2 > 0 { if clk_find(names, nn, nm2) < 0 { if nn < CLK_MAXJOBS { 472 var d2: *u8 = clk_slot(names, nn) 473 var a2: i64 = 0 474 while a2 < w2 { d2[a2] = nm2[a2]; a2 = a2 + 1 } 475 d2[w2] = 0 as u8 476 var e2: *u8 = clk_slot(organs, nn) 477 var b2: i64 = 0 478 while b2 < g2 { e2[b2] = og2[b2]; b2 = b2 + 1 } 479 e2[g2] = 0 as u8 480 if iv2 < 1 { iv2 = 1 } 481 intervals[nn] = iv2 482 next_due[nn] = nd2 483 nn = nn + 1 484 } } } } 485 } 486 ls = p + 1 487 } 488 p = p + 1 489 } 490 } 491 let buf: *u8 = sys_mmap(CLK_MAXJOBS*CLK_ROWW); var o: i64 = 0 492 var i: i64 = 0 493 while i < nn { 494 let nm: *u8 = clk_slot(names, i); var k: i64=0; while nm[k]!=(0 as u8){ buf[o]=nm[k]; o=o+1; k=k+1 } 495 buf[o]=9 as u8; o=o+1; o = clk_itoa(buf, o, intervals[i]); buf[o]=9 as u8; o=o+1; o = clk_itoa(buf, o, next_due[i]); buf[o]=9 as u8; o=o+1 496 let og: *u8 = clk_slot(organs, i); var g: i64=0; while og[g]!=(0 as u8){ buf[o]=og[g]; o=o+1; g=g+1 } 497 buf[o]=10 as u8; o=o+1 498 i = i + 1 499 } 500 sys_mkdir("knowledge" as *u8, 0x1ed); sys_mkdir("knowledge/sched" as *u8, 0x1ed) 501 let tmp: *u8 = sys_mmap(256); var t: i64=0; let pp: *u8 = path; while pp[t]!=(0 as u8){ tmp[t]=pp[t]; t=t+1 } tmp[t]=46 as u8; tmp[t+1]=116 as u8; tmp[t+2]=109 as u8; tmp[t+3]=112 as u8; tmp[t+4]=0 as u8 // path + ".tmp" 502 let fd: i64 = sys_openat_wr(tmp, 0x1a4); if fd<0 { return 0 } sys_write(fd, buf, o); sys_close(fd) 503 __syscall(82, tmp as i64, path as i64, 0, 0, 0, 0) // atomic rename .tmp -> registry 504 return nn 505} 506 507// ---- PLANE-NATIVE STATE PERSISTENCE (2026-08-03, debts 1784828927 + 1784868625: the operator law is 508// planes, never tsv). The mutable schedule state (name·interval·next_due·organ per row, the thing the 509// clock resumes from) lives in a seg-store plane, written once per window via sts_seed. The plane is 510// CLOCK-EXCLUSIVE by doctrine (external adds ride the separate clockjobs- ADD plane, which the clock 511// is read-only on) -- but the same reload-merge that saved the tsv from the 1785559048 clobber is kept 512// here: rows we have never seen are preserved with their own next_due, because A WRITER THAT 513// SERIALIZES ONLY ITS OWN MEMORY CANNOT COEXIST WITH ANY OTHER WRITER. -------------------------------- 514 515// load schedule state from the plane; same 4-col row grammar as the legacy file. returns #jobs. 516func clk_load_plane(prefix: *u8, names: *u8, organs: *u8, intervals: *i64, next_due: *i64, np: *i64) -> i64 { 517 np[0] = 0 518 let data: *u8 = sys_mmap(CLK_MAXJOBS*CLK_ROWW) 519 let dn: i64 = sts_load(prefix, data, CLK_MAXJOBS*CLK_ROWW) 520 if dn <= 0 { return 0 } 521 var i: i64=0; var ls: i64=0; var n: i64=0 522 while i < dn { 523 if data[i] == (10 as u8) { 524 if i > ls { if n < CLK_MAXJOBS { 525 let line: *u8 = ((data as i64)+ls) as *u8; let ll: i64 = i-ls 526 let d: *u8 = clk_slot(names, n); let e: *u8 = clk_slot(organs, n) 527 var f: i64=0; var p: i64=0; var iv: i64=0; var nd: i64=0; var w: i64=0; var g: i64=0 528 while p < ll { 529 if line[p] == (9 as u8) { f = f + 1 } 530 else { 531 if f == 0 { if w < CLK_NAMEW-1 { d[w]=line[p]; w=w+1 } } 532 else { if f == 3 { if g < CLK_NAMEW-1 { e[g]=line[p]; g=g+1 } } 533 else { if line[p] >= (48 as u8) { if line[p] <= (57 as u8) { 534 let dig: i64 = (line[p]-(48 as u8)) as i64 535 if f == 1 { iv = iv*10 + dig } else { nd = nd*10 + dig } 536 } } } } 537 } 538 p = p + 1 539 } 540 d[w] = 0 as u8; e[g] = 0 as u8 541 if w > 0 { if g > 0 { 542 intervals[n]=iv; next_due[n]=nd; n=n+1 543 } } 544 } } 545 ls = i + 1 546 } 547 i = i + 1 548 } 549 np[0] = n 550 // B-hunk (2026-08-22): this buffer is per-call and was never freed -- at CLK_ROWW=2112 that is a 551 // 1,081,344 B address-space leak per window on a daemon whose life is 120 windows. Everything it 552 // held was COPIED into the caller's slot arrays above, so the unmap is safe by construction. 553 sys_munmap(data, CLK_MAXJOBS*CLK_ROWW) 554 return n 555} 556 557// persist schedule state to the plane (reload-merge first, then ONE whole-plane sts_seed commit). 558// Rows WE hold win (ours carry the advanced next_due); unknown rows are preserved verbatim. 559func clk_save_plane(prefix: *u8, names: *u8, organs: *u8, intervals: *i64, next_due: *i64, n: i64) -> i64 { 560 var nn: i64 = n 561 let rnames: *u8 = sys_mmap(CLK_MAXJOBS*CLK_NAMEW) 562 let rorgs: *u8 = sys_mmap(CLK_MAXJOBS*CLK_NAMEW) 563 let riv: *i64 = sys_mmap(CLK_MAXJOBS*8) as *i64 564 let rnd: *i64 = sys_mmap(CLK_MAXJOBS*8) as *i64 565 let rnp: *i64 = sys_mmap(8) as *i64 566 clk_load_plane(prefix, rnames, rorgs, riv, rnd, rnp) 567 var ri: i64 = 0 568 while ri < rnp[0] { 569 if clk_find(names, nn, clk_slot(rnames, ri)) < 0 { if nn < CLK_MAXJOBS { 570 var d2: *u8 = clk_slot(names, nn); let sm: *u8 = clk_slot(rnames, ri) 571 var a2: i64 = 0 572 while a2 < CLK_NAMEW-1 { if sm[a2]==(0 as u8){ d2[a2]=0 as u8; a2=CLK_NAMEW } else { d2[a2]=sm[a2]; a2=a2+1 } } 573 if a2 == CLK_NAMEW-1 { d2[a2]=0 as u8 } 574 var e2: *u8 = clk_slot(organs, nn); let so: *u8 = clk_slot(rorgs, ri) 575 var b2: i64 = 0 576 while b2 < CLK_NAMEW-1 { if so[b2]==(0 as u8){ e2[b2]=0 as u8; b2=CLK_NAMEW } else { e2[b2]=so[b2]; b2=b2+1 } } 577 if b2 == CLK_NAMEW-1 { e2[b2]=0 as u8 } 578 intervals[nn] = riv[ri]; next_due[nn] = rnd[ri]; nn = nn + 1 579 } } 580 ri = ri + 1 581 } 582 let buf: *u8 = sys_mmap(CLK_MAXJOBS*CLK_ROWW); var o: i64 = 0 583 var i: i64 = 0 584 while i < nn { 585 let nm: *u8 = clk_slot(names, i); var k: i64=0; while nm[k]!=(0 as u8){ buf[o]=nm[k]; o=o+1; k=k+1 } 586 buf[o]=9 as u8; o=o+1; o = clk_itoa(buf, o, intervals[i]); buf[o]=9 as u8; o=o+1; o = clk_itoa(buf, o, next_due[i]); buf[o]=9 as u8; o=o+1 587 let og: *u8 = clk_slot(organs, i); var g: i64=0; while og[g]!=(0 as u8){ buf[o]=og[g]; o=o+1; g=g+1 } 588 buf[o]=10 as u8; o=o+1 589 i = i + 1 590 } 591 sts_seed(prefix, buf, o) 592 // B-hunk (2026-08-22): free the per-call buffers -- at the widened CLK_NAMEW/CLK_ROWW these 593 // mappings are ~3.2 MB per save (1-2 saves/window) of address space the GC-free substrate 594 // otherwise leaks until the 120-window life recycle. All contents are already committed/copied. 595 sys_munmap(rnames, CLK_MAXJOBS*CLK_NAMEW) 596 sys_munmap(rorgs, CLK_MAXJOBS*CLK_NAMEW) 597 sys_munmap(riv as *u8, CLK_MAXJOBS*8) 598 sys_munmap(rnd as *u8, CLK_MAXJOBS*8) 599 sys_munmap(rnp as *u8, 8) 600 sys_munmap(buf, CLK_MAXJOBS*CLK_ROWW) 601 return nn 602} 603 604// THE ONE STATE LOADER for the live clock: plane first; if the plane is empty AND a legacy tsv 605// exists, load it (the one-time migration path) -- the caller's next clk_save_plane completes the 606// cutover. Returns 0=loaded-from-plane, 1=migrated-from-legacy, -1=nothing anywhere. 607func clk_load_state(prefix: *u8, legacy: *u8, names: *u8, organs: *u8, intervals: *i64, next_due: *i64, np: *i64) -> i64 { 608 if clk_load_plane(prefix, names, organs, intervals, next_due, np) > 0 { return 0 } 609 if clk_load(legacy, names, organs, intervals, next_due, np) > 0 { return 1 } 610 return 0 - 1 611} 612 613// MERGE new job declarations from an nx_store PLANE (off-tsv, additive) into the in-memory registry. 614// THE CLOBBER FIX: the clock is READ-ONLY on this plane -- external adds go via 615// `nx_store_put <plane> put <actor> <name> <interval> <organ>` (upsert-by-name, additive) and can NEVER be 616// clobbered by the clock's own (now clock-private) tsv save. Plane row cols (TAB): 0=name 1=interval(sec) 2=organ. 617// A job already present (by name) is skipped (idempotent -- re-merge is a no-op). New jobs arm at base_tick+interval. 618// Returns the number of NEW jobs added this call. 619func clk_merge_store(prefix: *u8, names: *u8, organs: *u8, intervals: *i64, next_due: *i64, np: *i64, base_tick: i64) -> i64 { 620 let buf: *u8 = sys_mmap(CLK_MAXJOBS*CLK_ROWW) 621 let dn: i64 = sts_load(prefix, buf, CLK_MAXJOBS*CLK_ROWW) 622 if dn <= 0 { return 0 } 623 // HOISTED: allocating this inside the row loop would leak a page per row (the nx_ts_lumadiff scar). 624 let msg: *u8 = sys_mmap(CLK_ROWW) 625 // DESIRED interval per job index, 0 = not declared this pass. mmap returns zeroed pages. 626 let want: *i64 = sys_mmap(CLK_MAXJOBS*8) as *i64 627 // DESIRED ORGAN per job index (2026-08-07). This reconciler recorded only the desired INTERVAL, so a 628 // declaration could re-PERIOD a job but never re-TARGET it. MEASURED: I declared segguard 600s -> 629 // nx_segsweep.elf and the plane came back interval=600 organ=nx_segguard.sh -- the cadence moved and the 630 // command did not, which is strictly WORSE than either alone: it put the 717s shell script back on a 631 // 600s period it cannot meet. Caught only because I read the organ column back instead of trusting the 632 // interval that DID change. 633 // (STAR)A RECONCILER THAT SYNCS ONE FIELD OF A ROW WILL SILENTLY DESYNC THE OTHERS -- AND A PARTIAL 634 // RECONCILE IS MORE DANGEROUS THAN NONE, BECAUSE THE FIELD THAT MOVED PROVES IT WORKED. 635 let worg: *u8 = sys_mmap(CLK_MAXJOBS*CLK_NAMEW) 636 // B-hunk: ONE pair of row-scratch buffers for the whole scan (see the hoist note at `msg` above). 637 let nmrow: *u8 = sys_mmap(CLK_NAMEW) 638 let ogrow: *u8 = sys_mmap(CLK_NAMEW) 639 var added: i64 = 0 640 var reper: i64 = 0 641 var i: i64=0; var ls: i64=0 642 while i < dn { 643 if buf[i] == (10 as u8) { 644 if i > ls { 645 let line: *u8 = ((buf as i64)+ls) as *u8; let ll: i64 = i-ls 646 let nm: *u8 = nmrow; let og: *u8 = ogrow // B-hunk: hoisted buffers (were per-row mmaps = a page per desired row per window; the HOISTED note at `msg` above already names the class) 647 var f: i64=0; var p: i64=0; var iv: i64=0; var w: i64=0; var g: i64=0 648 while p < ll { 649 if line[p]==(9 as u8) { f=f+1 } 650 else { 651 if f==0 { if w<CLK_NAMEW-1 { nm[w]=line[p]; w=w+1 } } 652 else { if f==1 { if line[p]>=(48 as u8) { if line[p]<=(57 as u8) { iv=iv*10+((line[p]-(48 as u8)) as i64) } } } 653 else { if f==2 { if g<CLK_NAMEW-1 { og[g]=line[p]; g=g+1 } } } } 654 } 655 p=p+1 656 } 657 nm[w]=0 as u8; og[g]=0 as u8 658 // MALFORMED-ROW GUARD (2026-08-03, debt 1784604739): a row lacking a name or an organ 659 // (e.g. a put that dropped an arg) must be SEEN, never silently skipped -- but never 660 // registered either (a nameless job / organless exec is garbage in the dispatcher). 661 if w == 0 { sts_werr("clk_merge_store: MALFORMED plane row (empty name) SKIPPED\n" as *u8) } 662 if w > 0 { if g == 0 { sts_werr("clk_merge_store: MALFORMED plane row (no organ) SKIPPED\n" as *u8) } } 663 if w > 0 { if g > 0 { 664 if iv < 1 { iv = 1 } 665 let n: i64 = np[0] 666 // ---- RECONCILE, NOT JUST ADD (2026-08-06) ---------------------------------------- 667 // MEASURED DEFECT: this merge was ADD-ONLY -- a row whose name already existed was 668 // silently skipped -- and clocksched- (the runtime state) is rewritten by the clock 669 // every window. So there was NO SUPPORTED PATH TO RE-PERIOD OR REMOVE A CLOCK JOB: 670 // you could add one and never change it again. That is not a hypothetical. Three jobs 671 // were stuck at pathological cadences -- bootstrap-gate at 60s and worldgen-gate at 672 // 120s are GATES (test suites) running 60-360x more often than every other gate in the 673 // plane (p384kat/tls12prf 21600, uigensitegate 86400), and together with intakekeep 674 // they are 60 of the ~273 dispatches/window = 22pct of ALL demand on a clock measured 675 // at 51 completed vs 273 demanded (5.4x oversubscribed, 31 jobs STARVED per window). 676 // Whoever added them could not fix them, and the starvation was blamed on scheduling 677 // policy for days while EDF was working exactly as designed. 678 // THE FIX IS THE ESTATE'S OWN PATTERN, USED TWICE ALREADY: cron.reg -> crontab via 679 // nx_cron_reconcile, and edge443 -> iptables via nx_edge443_reconcile.sh. The declared 680 // registry is the SSOT and the runtime state is RECONCILED toward it every window. 681 // clockjobs- is now DESIRED STATE, not an append-only inbox. 682 // SAFETY: re-arm to base_tick+iv rather than leaving the old deadline, so a re-period 683 // can never fire a herd (shortening) nor strand a job past its new period (lengthening). 684 // ★A REGISTRY YOU CAN ONLY APPEND TO IS A REGISTRY THAT ACCUMULATES ITS OWN MISTAKES. 685 let ex: i64 = clk_find(names, n, nm) 686 // RECORD the intent; do not apply it here. See the apply loop after this scan. 687 if ex >= 0 { want[ex] = iv 688 var wd: *u8 = clk_slot(worg, ex); var wk: i64 = 0 689 while wk < CLK_NAMEW-1 { if og[wk]==(0 as u8){wd[wk]=0 as u8;wk=CLK_NAMEW} else {wd[wk]=og[wk];wk=wk+1} } 690 if wk == CLK_NAMEW-1 { wd[wk] = 0 as u8 } 691 } 692 if ex < 0 { if n < CLK_MAXJOBS { 693 var d: *u8=clk_slot(names,n); var k: i64=0; while k<CLK_NAMEW-1 { if nm[k]==(0 as u8){d[k]=0 as u8;k=CLK_NAMEW} else {d[k]=nm[k];k=k+1} } if k==CLK_NAMEW-1 {d[k]=0 as u8} 694 var e: *u8=clk_slot(organs,n); var j: i64=0; while j<CLK_NAMEW-1 { if og[j]==(0 as u8){e[j]=0 as u8;j=CLK_NAMEW} else {e[j]=og[j];j=j+1} } if j==CLK_NAMEW-1 {e[j]=0 as u8} 695 if iv<1 { iv=1 } 696 intervals[n]=iv; next_due[n]=base_tick+iv; np[0]=n+1; added=added+1 697 } } 698 } } 699 } 700 ls=i+1 701 } 702 i=i+1 703 } 704 // APPLY ONCE PER JOB, AFTER EVERY ROW IS READ -- LAST DECLARATION WINS. 705 // Applying inside the row loop was the first cut and it was wrong: this plane ACCUMULATES rows 706 // (pubreconcile appears THREE times in it today), so a name carrying two differing declarations 707 // would flip-flop -- two writes and two log lines every window, forever, converging on exactly 708 // the same last-row-wins answer this loop reaches quietly. Collecting the intent and applying it 709 // once is both quieter and identical in outcome. 710 var wi: i64 = 0 711 while wi < np[0] { 712 if want[wi] > 0 { if intervals[wi] != want[wi] { 713 var mo: i64 = clk_msgcat(msg, 0, "clk_merge_store: REPERIOD " as *u8) 714 mo = clk_msgcat(msg, mo, clk_slot(names, wi)) 715 mo = clk_msgcat(msg, mo, " " as *u8) 716 mo = clk_itoa(msg, mo, intervals[wi]) 717 mo = clk_msgcat(msg, mo, "s -> " as *u8) 718 mo = clk_itoa(msg, mo, want[wi]) 719 mo = clk_msgcat(msg, mo, "s (declared in clockjobs-; re-armed at now+interval)\n" as *u8) 720 msg[mo] = 0 as u8 721 sts_werr(msg) 722 intervals[wi] = want[wi] 723 next_due[wi] = base_tick + want[wi] 724 reper = reper + 1 725 } } 726 // RE-TARGET: the organ is a field of the row too, and a declaration that names a different 727 // binary must MOVE the job, not just its cadence. Separate from the interval branch so a 728 // pure re-target (same period, new organ) is honoured. 729 if wi < np[0] { let wo: *u8 = clk_slot(worg, wi) 730 if wo[0] != (0 as u8) { if clk_streq(wo, clk_slot(organs, wi)) == 0 { 731 var mo2: i64 = clk_msgcat(msg, 0, "clk_merge_store: RETARGET " as *u8) 732 mo2 = clk_msgcat(msg, mo2, clk_slot(names, wi)) 733 mo2 = clk_msgcat(msg, mo2, " -> " as *u8) 734 mo2 = clk_msgcat(msg, mo2, wo) 735 mo2 = clk_msgcat(msg, mo2, " (declared in clockjobs-)\n" as *u8) 736 msg[mo2] = 0 as u8 737 sts_werr(msg) 738 var od: *u8 = clk_slot(organs, wi); var ok2: i64 = 0 739 while ok2 < CLK_NAMEW-1 { if wo[ok2]==(0 as u8){od[ok2]=0 as u8;ok2=CLK_NAMEW} else {od[ok2]=wo[ok2];ok2=ok2+1} } 740 if ok2 == CLK_NAMEW-1 { od[ok2] = 0 as u8 } 741 reper = reper + 1 742 } } } 743 wi = wi + 1 744 } 745 // RETURN ADDS *AND* RE-PERIODS (2026-08-06). This returned only `added`, so a caller could not tell 746 // that a RECONCILE had occurred -- reper was counted, logged and then dropped on the floor. The single 747 // caller discarded the value entirely, so widening it is safe and makes the reconcile actionable. 748 // (STAR)A COUNTER THAT IS COMPUTED, LOGGED AND NOT RETURNED IS A FACT THE CALLER CANNOT ACT ON. 749 // B-hunk: free the per-call transients (buf/msg/want/worg + the hoisted row scratch); everything 750 // they held was applied to the caller's arrays above. 751 sys_munmap(buf, CLK_MAXJOBS*CLK_ROWW) 752 sys_munmap(msg, CLK_ROWW) 753 sys_munmap(want as *u8, CLK_MAXJOBS*8) 754 sys_munmap(worg, CLK_MAXJOBS*CLK_NAMEW) 755 sys_munmap(nmrow, CLK_NAMEW) 756 sys_munmap(ogrow, CLK_NAMEW) 757 return added + reper 758} 759 760// ============================================================================================ 761// EDF / WALL-CLOCK SCHEDULING (2026-08-04, debt 1785872141 -- operator: "get the clock to sota") 762// 763// THE MEASURED DEFECT clk_run_tickless has BY CONSTRUCTION: its logical tick T advances ONLY by the 764// sleep amount (T = minnd), while REAL time also advances by however long the dispatched children 765// took (clk_dispatch_run forks and sys_wait4s EVERY job, serially, inside the beat). So logical time 766// drifts behind wall-clock time in proportion to dispatch load, and EVERY job's period silently 767// STRETCHES: measured 2026-08-04, evidencebeat (interval 21600 = "6 hours") fired ONCE in 21 HOURS 768// while the clock was demonstrably alive (tick current, window-end a=30 b=132). Even 60s netobs 769// slipped to 243s. The fast jobs pay the drift too, but a 6h job pays it 360x over. 770// u2605u2605u2605u2605u2605u2605 A SCHEDULER THAT ADVANCES ITS OWN CLOCK BY WHAT IT SLEPT -- NOT BY WHAT ELAPSED -- 771// MEASURES ITS OWN IDLENESS AND CALLS IT TIME. Every deadline it derives is then a lie under load. 772// 773// THE SOTA SHAPE (what real schedulers do; cron/systemd-timers/EDF literature all agree): 774// 1. deadlines are ABSOLUTE WALL-CLOCK instants, never a self-advanced counter 775// 2. re-read the clock AFTER every dispatch, so job runtime cannot be lost 776// 3. dispatch EARLIEST-DEADLINE-FIRST so the most-overdue job goes first (EDF is optimal for 777// meeting deadlines on one resource -- and it is exactly what stops slow-job starvation) 778// 4. catch-up without runaway: re-arm past now by WHOLE intervals, and REPORT missed periods 779// instead of pretending they happened 780// The decision math is factored into PURE functions below precisely so a gate can prove a 6-hour 781// period behaves correctly in MILLISECONDS with fabricated clock values -- a scheduler you can only 782// test by waiting 6 hours is a scheduler nobody tests. 783 784// deadlines below this are LEGACY LOGICAL TICKS (the old counter ran ~3e6; epochs are ~1.78e9), so the 785// one-time migration is unambiguous and needs no flag day. NEVER compare an epoch against a tick. 786const CLK_EPOCH_FLOOR: i64 = 1000000000 787 788// PURE: index of the overdue job with the EARLIEST deadline (EDF), or -1 if nothing is due. 789// Ties break toward the lower index = stable, so a tie can never rotate two jobs into each other's slot. 790// DEMAND vs CAPACITY (2026-08-06, measured on the live plane). EDF decides WHO runs next; it can never 791// decide HOW MANY CAN RUN. If the registry asks for more dispatches per window than the window can 792// complete, then EVERY ordering starves someone and the scheduler is merely choosing the victim -- so 793// the deficit has to be a NUMBER the clock PUBLISHES, not something inferred from a late heartbeat days 794// later. MEASURED THE DAY THIS SHIPPED: 45 jobs demanded ~273 dispatches per 1800s window while the 795// window completed 51 (~35s of real child time per dispatch) = ~5.4x oversubscribed, and 273 exceeds even 796// TLN_MAXDISPATCH=240, so the set could not be served at INFINITE speed. That is why netobs (60s period) 797// was firing at 284s+ intervals with its registry row and its elf both perfectly healthy. 798// CONSERVATIVE BY CONSTRUCTION: integer division FLOORS, and any job whose period exceeds the window 799// contributes 0. So this is a LOWER BOUND on demand -- if even this floor exceeds capacity, 800// oversubscription is PROVEN, never merely suspected. 801func clk_demand_per_window(intervals: *i64, n: i64, window_secs: i64) -> i64 { 802 var need: i64 = 0 803 var i: i64 = 0 804 while i < n { 805 if intervals[i] > 0 { need = need + (window_secs / intervals[i]) } 806 i = i + 1 807 } 808 return need 809} 810 811// PARALLEL-SAFE PICK (2026-08-06): identical to clk_edf_pick except it SKIPS jobs already IN FLIGHT. 812// THE HAZARD IT EXISTS FOR, and it is not hypothetical: a job is re-armed only AFTER its child is reaped, 813// so while a run is in progress its deadline is still in the past -- and a job whose runtime EXCEEDS its 814// period (worldgen: 289s measured against a 120s declared period) is overdue for the WHOLE of its own run. 815// Under bounded-parallel dispatch a picker without this mask would immediately dispatch a SECOND copy of 816// the exact organ already saturating the box, then a third, up to the slot cap. Two concurrent four-world 817// renders is strictly WORSE than the serial starvation parallelism was introduced to fix. 818// (STAR)PARALLELISM TURNS "ALWAYS OVERDUE" FROM A STARVATION BUG INTO A FORK BOMB -- THE BUSY MASK IS NOT 819// AN OPTIMISATION, IT IS THE CORRECTNESS CONDITION. Kept as a SEPARATE pure function so it is gate-provable 820// on fabricated inputs (T10/T11) before it is ever wired to a live dispatcher. 821func clk_edf_pick_free(deadlines: *i64, n: i64, now: i64, busy: *i64) -> i64 { 822 var best: i64 = 0 - 1 823 var bestd: i64 = 0 824 var i: i64 = 0 825 while i < n { 826 if busy[i] == 0 { 827 if deadlines[i] <= now { 828 if best < 0 { best = i; bestd = deadlines[i] } 829 else { if deadlines[i] < bestd { best = i; bestd = deadlines[i] } } 830 } 831 } 832 i = i + 1 833 } 834 return best 835} 836 837func clk_edf_pick(deadlines: *i64, n: i64, now: i64) -> i64 { 838 var best: i64 = 0 - 1 839 var bestd: i64 = 0 840 var i: i64 = 0 841 while i < n { 842 if deadlines[i] <= now { 843 if best < 0 { best = i; bestd = deadlines[i] } 844 else { if deadlines[i] < bestd { best = i; bestd = deadlines[i] } } 845 } 846 i = i + 1 847 } 848 return best 849} 850 851// PURE: soonest deadline across all jobs (what a tickless sleeper must sleep until). -1 if n==0. 852func clk_edf_next(deadlines: *i64, n: i64) -> i64 { 853 if n <= 0 { return 0 - 1 } 854 var m: i64 = deadlines[0] 855 var i: i64 = 1 856 while i < n { if deadlines[i] < m { m = deadlines[i] } i = i + 1 } 857 return m 858} 859 860// PURE: how many whole periods job i has MISSED at `now` (0 = on time / early). This is the honest 861// starvation measure -- a job 3 periods late is a 3, not a "fired". 862func clk_edf_missed(deadlines: *i64, intervals: *i64, i: i64, now: i64) -> i64 { 863 if deadlines[i] > now { return 0 } 864 var iv: i64 = intervals[i] 865 if iv < 1 { iv = 1 } 866 return (now - deadlines[i]) / iv 867} 868 869// PURE: re-arm job i past `now` by WHOLE intervals (catch-up safe: never fires twice for one missed 870// window, never drifts off-phase). Returns the periods skipped so the caller can REPORT them. 871func clk_edf_rearm(deadlines: *i64, intervals: *i64, i: i64, now: i64) -> i64 { 872 var iv: i64 = intervals[i] 873 if iv < 1 { iv = 1 } 874 var skipped: i64 = 0 875 while deadlines[i] <= now { deadlines[i] = deadlines[i] + iv; skipped = skipped + 1 } 876 if skipped > 0 { skipped = skipped - 1 } // the first advance is the fire itself, not a miss 877 return skipped 878} 879 880// PURE: one-time migration of legacy logical-tick deadlines to wall-clock instants. A tick value can 881// never be a valid epoch, so this is exact. Jobs arm at now+interval (their author's intent) rather 882// than firing a thundering herd at now. Returns how many rows were converted. 883func clk_edf_migrate(deadlines: *i64, intervals: *i64, n: i64, now: i64) -> i64 { 884 var c: i64 = 0 885 var i: i64 = 0 886 while i < n { 887 if deadlines[i] < CLK_EPOCH_FLOOR { 888 var iv: i64 = intervals[i] 889 if iv < 1 { iv = 1 } 890 deadlines[i] = now + iv 891 c = c + 1 892 } 893 i = i + 1 894 } 895 return c 896} 897 898// THE EDF WINDOW: wall-clock anchored, EDF-ordered, starvation-reporting. Runs until `maxdispatch` 899// jobs have been dispatched or `budget_secs` of REAL time is gone, sleeping only when nothing is due. 900// out[0]=dispatched out[1]=slept_secs out[2]=final_now out[3]=starved_jobs out[4]=max_lateness_secs 901// out[5]=exec_secs (real time inside children -- the number the old design silently threw away). 902// ---- IN-FLIGHT BEAT (2026-08-06) ------------------------------------------------------------- 903// PROVEN LIVE THIS SESSION, NOT HYPOTHESISED: the clock overran its own 1800s budget by 42s and 904// stopped dispatching for 475s+ while one child held it, because clk_dispatch_one waits with NO 905// deadline (debt 1786056459). sitecheck (30s period) froze; every beat in the estate went dead. 906// THE INSTRUMENT PROBLEM THAT INCIDENT EXPOSED: max_exec, starved, and the entire window-end report 907// are written AFTER clk_run_edf RETURNS. So the one failure mode that prevents returning is exactly 908// the one the report can never describe -- during the stall the newest window-end beat was 1800s old 909// and named a completely different window. 910// (STAR)AN INSTRUMENT THAT REPORTS ONLY AT COMPLETION CANNOT REPORT WHAT PREVENTS COMPLETION. 911// This beat is written BEFORE every dispatch and overwritten AFTER it, so a hang is visible the 912// moment it starts and NAMES the organ holding the scheduler. Cost is 2 file writes per dispatch, 913// negligible beside the fork+execve it brackets. 914func clk_ifw(s: *u8, dst: *u8, o: i64) -> i64 { var x: i64=o; var i: i64=0; while s[i]!=(0 as u8){dst[x]=s[i];x=x+1;i=i+1} return x } 915func clk_ifn(dst: *u8, o: i64, v: i64) -> i64 { var x: i64=o; var mm: i64=v; if mm<0{dst[x]=45 as u8;x=x+1;mm=0-mm} if mm==0{dst[x]=48 as u8;return x+1} let t:*u8=sys_mmap(24); var k:i64=0; while mm>0{t[k]=(48+(mm%10)) as u8;mm=mm/10;k=k+1} var j:i64=0; while j<k{dst[x]=t[k-1-j];x=x+1;j=j+1} return x } 916func clk_inflight(tag: *u8, organ: *u8, a: i64, b: i64) -> i64 { 917 let st: *u8 = sys_mmap(CLK_NAMEW + 128); var o: i64 = 0 // composes the ORGAN slot + <=97 fixed bytes; DERIVED from the slot (1024 was safe only by mmap page rounding once the slot widened) 918 o = clk_ifw("CLOCKBEAT " as *u8, st, o); o = clk_ifw(tag, st, o) 919 o = clk_ifw(" organ=" as *u8, st, o); o = clk_ifw(organ, st, o) 920 o = clk_ifw(" a=" as *u8, st, o); o = clk_ifn(st, o, a) 921 o = clk_ifw(" b=" as *u8, st, o); o = clk_ifn(st, o, b) 922 o = clk_ifw(" t=" as *u8, st, o); o = clk_ifn(st, o, sys_now_realtime_sec()) 923 st[o]=10 as u8; o=o+1; st[o]=0 as u8 924 let fd: i64 = sys_openat_wr("sites/nishifamily/clock_inflight.txt" as *u8, 0x1a4) 925 if fd >= 0 { sys_write(fd, st, o); sys_close(fd) } 926 return 0 927} 928 929func clk_run_edf(organs: *u8, names: *u8, intervals: *i64, deadlines: *i64, n: i64, maxdispatch: i64, budget_secs: i64, tick_ms: i64, out: *i64) -> i64 { 930 let fired: *i64 = sys_mmap(CLK_MAXJOBS*8) as *i64 931 let t0: i64 = sys_now_realtime_sec() 932 var disp: i64 = 0 933 var slept: i64 = 0 934 var starved: i64 = 0 935 var maxlate: i64 = 0 936 var execs: i64 = 0 937 var maxexec: i64 = 0 938 var maxidx: i64 = 0 - 1 939 // STARVED AXIS, NAMED (2026-08-14). Mirrors maxexec/maxidx below, which already record the worst 940 // single dispatch AND which job owned it. That law was applied to the SLOW axis and never to the 941 // STARVED axis, so out[3] shipped as a bare count. 942 var maxmissed: i64 = 0 943 var starvidx: i64 = 0 - 1 944 // STORM-AWARE DISPATCH state (2026-09-02): the heavy list is read ONCE per window (sys_read_file sizes 945 // from the file, cannot short-read; an absent conf -> hn=0 -> nothing is heavy), the ruler's scratch is 946 // allocated ONCE, and the deferral count is published in out[10]. 947 var deferred: i64 = 0 948 let ioa: *i64 = sys_mmap(16) as *i64 949 let hlen: *i64 = sys_mmap(16) as *i64 950 hlen[0] = 0 951 let hconf: *u8 = sys_read_file(CLK_HEAVY_CONF, hlen) 952 var hn: i64 = 0 953 if (hconf as i64) != 0 { hn = hlen[0] } 954 var run: i64 = 1 955 while run == 1 { 956 // RE-READ THE CLOCK EVERY ITERATION -- this single line is the fix: child runtime is now 957 // observed, not assumed away. 958 let now: i64 = sys_now_realtime_sec() 959 if now - t0 >= budget_secs { run = 0 } 960 if disp >= maxdispatch { run = 0 } 961 if run == 1 { 962 var k: i64 = clk_edf_pick(deadlines, n, now) 963 // STORM-AWARE DISPATCH (2026-09-02): a HEAVY job (listed in CLK_HEAVY_CONF) that is due while the 964 // shared I/O-storm ruler reports zero spawn budget is re-armed to now + clk_defer_secs(interval) 965 // and NOT forked this pass; k becomes -1 so the else-branch below sleeps until the next real 966 // deadline exactly as if nothing were due. Light jobs, an absent conf and an UNREADABLE ruler 967 // take the pre-change path byte-for-byte. The deferral is announced on the in-flight beat with 968 // b = procs_blocked, so the docroot witness names the storm that caused it. 969 if k >= 0 { if clk_heavy_listed(hconf, hn, clk_slot(names, k)) == 1 { 970 let bud: i64 = clk_storm_budget(ioa) 971 if clk_storm_defer(1, bud) == 1 { 972 deferred = deferred + 1 973 deadlines[k] = now + clk_defer_secs(intervals[k]) 974 clk_inflight("deferred-storm" as *u8, clk_slot(organs, k), ioa[1], k) 975 k = 0 - 1 976 } 977 } } 978 if k >= 0 { 979 let missed: i64 = clk_edf_missed(deadlines, intervals, k, now) 980 let late: i64 = now - deadlines[k] 981 if late > maxlate { maxlate = late } 982 if missed > 0 { 983 starved = starved + 1 984 // NAME THE STARVER, DO NOT MERELY COUNT IT. A window reported "STARVED 3" and nothing 985 // anywhere said WHICH 3, so a job persistently losing EDF was undiagnosable. 986 // MEASURED 2026-08-14: a window published b=3 STARVED while nx_resmon's leak axis was 987 // flipping to UNOBSERVABLE off a stale memvel snapshot -- and there was no way to tell 988 // whether memvel was one of the three. A COUNT WITHOUT A WORKLIST IS NOT ACTIONABLE. 989 // Worst-offender (most periods missed) rather than a list: it needs no allocation, it 990 // matches the maxexec/maxidx idiom already proven here, and the worst starver is the 991 // one whose period is actually unschedulable. 992 if missed > maxmissed { maxmissed = missed; starvidx = k } 993 } 994 clk_edf_rearm(deadlines, intervals, k, now) 995 let e0: i64 = now 996 clk_inflight("in-flight" as *u8, clk_slot(organs, k), e0, k) 997 let ran: i64 = clk_dispatch_one(organs, names, k) 998 let e1: i64 = sys_now_realtime_sec() 999 clk_inflight("idle" as *u8, clk_slot(organs, k), e1 - e0, k) 1000 // ---- COST-ADAPTIVE BACKOFF (2026-08-06) -------------------------------------- 1001 // PROVEN LIVE, NAMED BY THE IN-FLIGHT BEAT ABOVE: nx_worldgen_gate.sov.elf holds a 1002 // 120s-period slot for 500s+ per run. A job whose RUNTIME EXCEEDS ITS PERIOD is always 1003 // overdue, so EDF correctly picks it again immediately, and on a SERIAL dispatcher with 1004 // no deadline that is a PERMANENT DENIAL OF SERVICE to every other job -- sitecheck (30s) 1005 // and netobs (60s) both froze behind it and every beat in the estate went dead. 1006 // (STAR)A JOB WHOSE RUNTIME EXCEEDS ITS PERIOD CANNOT BE SCHEDULED AT THAT PERIOD -- THE 1007 // REGISTRY IS ASSERTING A COST NOBODY EVER MEASURED. So MEASURE IT AND BELIEVE THE 1008 // MEASUREMENT: re-arm such a job to now + its OWN observed runtime instead of now + 1009 // its declared interval. It still runs, it still makes progress, but it can never again 1010 // starve the whole registry, and the declared interval becomes a FLOOR not a fiction. 1011 // Chosen over SIGKILL deliberately: killing needs a raw signal syscall inside a forking 1012 // scheduler, and this tree has a live trap where literal syscall numbers are rv64->x86 1013 // translated. This fix needs NO new syscall and cannot mis-target a pid. 1014 if e1 - e0 > intervals[k] { deadlines[k] = e1 + (e1 - e0) * CLK_HOGDUTY } // dt is declared BELOW this point; use the operands already in scope 1015 let dt: i64 = e1 - e0 1016 execs = execs + dt 1017 // ★PUBLISH A MAX BEFORE YOU ACT ON A MEAN (2026-08-06). Aggregate exec_secs cannot separate 1018 // 'every child is a bit slow' from 'ONE child hung for most of the window' -- both produce the 1019 // same dispatch count and the same mean, and they have OPPOSITE remedies (raise capacity vs 1020 // kill one hog). clk_dispatch_one waits on the child with NO deadline (debt 1786056459), so the 1021 // hung case is not hypothetical. Record the worst SINGLE dispatch and WHICH job owned it. 1022 if dt > maxexec { maxexec = dt; maxidx = k } 1023 if ran == 1 { disp = disp + 1; fired[k] = 1 } 1024 } else { 1025 let nd: i64 = clk_edf_next(deadlines, n) 1026 if nd < 0 { run = 0 } else { 1027 var wait: i64 = nd - now 1028 if wait < 1 { wait = 1 } 1029 let left: i64 = budget_secs - (now - t0) 1030 if wait > left { wait = left } 1031 if wait < 1 { run = 0 } else { 1032 let ts: *i64 = sys_mmap(16) as *i64 1033 ts[0] = wait 1034 ts[1] = 0 1035 __syscall(35, ts as i64, 0, 0, 0, 0, 0) 1036 slept = slept + wait 1037 } 1038 } 1039 } 1040 } 1041 } 1042 out[0] = disp; out[1] = slept; out[2] = sys_now_realtime_sec() 1043 out[3] = starved; out[4] = maxlate; out[5] = execs 1044 out[6] = maxexec; out[7] = maxidx // worst single dispatch, and the job index that owned it 1045 out[8] = starvidx; out[9] = maxmissed // WORST STARVER: job index, and how many of its periods it missed 1046 out[10] = deferred // STORM-DEFERRED (2026-09-02): heavy dispatches this window declined on the shared I/O-storm ruler; the caller allocates 16 slots (128 bytes), so slot 10 is inside the mapping 1047 // CALLER CONTRACT: `out` must now be at least 10 i64 slots (80 bytes). nx_clock_tickless allocated 1048 // exactly 64 bytes = 8 slots, so writing out[8] there would have run off the end of the mapping -- 1049 // the caller was widened in the same change. Any new caller must allocate >= 80 bytes. 1050 return disp 1051} 1052 1053// THE ARGV SPLIT, AS A PURE FUNCTION (2026-08-22, command-slot truncation fix). Copies the organ 1054// slot into `buf` (>= CLK_NAMEW+8 bytes), splits on single spaces IN PLACE, fills `argv` 1055// (>= 8*(CLK_ARGV_MAX+1) bytes) and returns the token count; argv[ac] is NULL. Factored out of 1056// clk_dispatch_one for the same reason the EDF math is factored into pure functions above: so a gate 1057// can prove an 18-token, 300-byte command survives the split WITHOUT a fork, and so the mutation 1058// site (CLK_NAMEW) is reachable from a tooth. 1059// OVERFLOW IS ANNOUNCED, NEVER SILENT: a command carrying more than CLK_ARGV_MAX tokens keeps the 1060// first CLK_ARGV_MAX and reports ARGV-OVERFLOW naming the job and the dropped count -- the old 1061// `if ac < 16` cap dropped every token past 15 with NO diagnostic, the same silent class as the 1062// 127-byte cut this change retires. Token-accept logic is otherwise byte-identical to the old code. 1063func clk_split_argv(raw0: *u8, buf: *u8, argv: *i64, label: *u8) -> i64 { 1064 var bl: i64 = 0 1065 var cdone: i64 = 0 1066 while cdone == 0 { 1067 if bl >= CLK_NAMEW - 1 { cdone = 1 } 1068 else { if raw0[bl] == (0 as u8) { cdone = 1 } else { buf[bl] = raw0[bl]; bl = bl + 1 } } 1069 } 1070 buf[bl] = 0 as u8 1071 argv[0] = buf as i64 1072 var ac: i64 = 1 1073 var dropped: i64 = 0 1074 var sp: i64 = 0 1075 while sp < bl { 1076 if buf[sp] == (32 as u8) { 1077 buf[sp] = 0 as u8 1078 if buf[sp+1] != (0 as u8) { 1079 if ac < CLK_ARGV_MAX { argv[ac] = ((buf as i64) + sp + 1); ac = ac + 1 } else { dropped = dropped + 1 } 1080 } 1081 } 1082 sp = sp + 1 1083 } 1084 argv[ac] = 0 1085 if dropped > 0 { 1086 let om: *u8 = sys_mmap(CLK_ROWW) 1087 var oo: i64 = clk_msgcat(om, 0, "clk_split_argv: ARGV-OVERFLOW job=" as *u8) 1088 oo = clk_msgcat(om, oo, label) 1089 oo = clk_msgcat(om, oo, " kept=" as *u8) 1090 oo = clk_itoa(om, oo, ac) 1091 oo = clk_msgcat(om, oo, " dropped=" as *u8) 1092 oo = clk_itoa(om, oo, dropped) 1093 oo = clk_msgcat(om, oo, " tail arguments NOT passed (CLK_ARGV_MAX in nx_clock_caps.nx bounds the vector; the bound is NAMED and this line is its announcement)\n" as *u8) 1094 om[oo] = 0 as u8 1095 sts_werr(om) 1096 sys_munmap(om, CLK_ROWW) 1097 } 1098 return ac 1099} 1100 1101// dispatch EXACTLY ONE job by index (the EDF loop picks the victim; this just runs it). Split out of 1102// clk_dispatch_run so ordering policy and exec mechanics are separable -- and so the EDF loop can 1103// re-read the clock between jobs. Returns 1 if the organ really ran, 0 if exec failed (127). 1104func clk_dispatch_one(organs: *u8, names: *u8, i: i64) -> i64 { 1105 // ARGV SUPPORT (2026-08-07). This built argv[0]=path, argv[1]=0 -- an organ could take NO ARGUMENTS. 1106 // (STAR)A DISPATCHER THAT CANNOT PASS ARGUMENTS MANUFACTURES SHELL SCRIPTS -- THE WRAPPERS ARE A 1107 // SYMPTOM OF THE CALLING CONVENTION, NOT OF LAZINESS. 1108 // The organ field is SPLIT ON SPACES into a real argv, so a row may read 1109 // fallbackharden<TAB>600<TAB>nx_actlog.elf harden knowledge/status/fallback.jrnl 1110 // BACK-COMPATIBLE BY CONSTRUCTION: a field with no space yields argv[0] only, and the slot is 1111 // CLK_NAMEW = CLK_CMD_CAP (nx_clock_caps.nx, shared with the writer nx_clockjob), so a command the 1112 // writer admits ALWAYS fits. The split itself is clk_split_argv above -- pure, gate-provable, 1113 // and it announces ARGV-OVERFLOW instead of silently dropping tokens. 1114 let buf: *u8 = sys_mmap(CLK_NAMEW + 8) 1115 let argv: *i64 = sys_mmap(8*(CLK_ARGV_MAX+1)) as *i64 1116 clk_split_argv(clk_slot(organs, i), buf, argv, clk_slot(names, i)) 1117 let path: *u8 = buf 1118 let path: *u8 = buf 1119 __syscall(90, path as i64, 0x1ed, 0, 0, 0, 0) // chmod +x the EXECUTABLE only, after the split 1120 let pid: i64 = sys_fork() 1121 if pid == 0 { 1122 // PER-JOB OUTPUT CAPTURE (2026-08-07). The SECOND service the .sh wrappers rendered, after argv: 1123 // nx_fallback_harden.cron.sh appends to logs/fallback_harden.log and its own header calls that 1124 // "the trend IS the telemetry". Replacing a wrapper with a bare organ row would have SILENTLY 1125 // LOST that log -- stripping a feature (rule 25) to remove a shell script. 1126 // (STAR)BEFORE REPLACING A WRAPPER, ENUMERATE EVERYTHING IT PROVIDES -- THE ARGUMENT YOU NOTICED 1127 // IS RARELY THE ONLY SERVICE IT RENDERS. 1128 // Every clock organ now gets, for free, the capture each wrapper hand-rolled: stdout AND stderr 1129 // append to logs/<jobname>.log. Named by JOB, not organ, so two jobs sharing a binary stay 1130 // distinguishable (clobbertest and frontdoor both run nx_atlas_frontdoor.elf). Fail-open: if the 1131 // log cannot be opened the child still execs -- telemetry must never be able to stop the beat. 1132 // Uses the sys_dup3 redirect idiom already proven in nx_guarded_run. 1133 let lp: *u8 = sys_mmap(CLK_NAMEW + 16); var lo: i64 = 0 // "logs/" + NAME slot + ".log" + NUL; DERIVED from the slot 1134 lo = clk_msgcat(lp, lo, "logs/" as *u8) 1135 lo = clk_msgcat(lp, lo, clk_slot(names, i)) 1136 lo = clk_msgcat(lp, lo, ".log" as *u8) 1137 lp[lo] = 0 as u8 1138 let ofd: i64 = sys_openat_append(lp, 0x1a4) 1139 if ofd >= 0 { sys_dup3(ofd, 1, 0); sys_dup3(ofd, 2, 0) } 1140 let envp: *i64 = sys_mmap(16) as *i64; envp[0] = "PATH=/usr/bin:/bin" as *u8 as i64; envp[1] = 0 1141 sys_execve(path, argv, envp) 1142 sys_exit(127) 1143 } 1144 let st: *i64 = sys_mmap(16) as *i64; st[0] = 0 1145 // BOUNDED WAIT (was a bare blocking wait4 -- the single line that let one child own the clock). 1146 let t0ms: i64 = sys_now_ms() 1147 var reaped: i64 = 0 1148 var timedout: i64 = 0 1149 while reaped == 0 { 1150 let w: i64 = sys_wait4(pid, st, 1) 1151 if w == pid { reaped = 1 } 1152 else { if w < 0 { reaped = 1 } 1153 else { if sys_now_ms() - t0ms >= CLK_DISPATCH_DEADLINE_MS { nx_kill(pid, CLK_SIGKILL); sys_wait4(pid, st, 0); reaped = 1; timedout = 1 } else { sys_sleep_ms(CLK_DISPATCH_POLL_MS) } } } 1154 } 1155 // A REAPED-ON-DEADLINE CHILD DID EXECUTE -- report it as dispatched so the duty ceiling above backs 1156 // it off from its OWN measured cost. Reporting 0 here would hide the hog from the very rule that tames it. 1157 // LIVE-PATH TELEMETRY (2026-08-07). clk_actlog was first added to clk_dispatch_run -- which is the 1158 // DEAD path: nx_clock_tickless calls clk_run_edf (it REPLACES clk_run_tickless), and clk_run_edf 1159 // dispatches through THIS function. The instrumented binary was verifiably the one running, and it 1160 // still logged nothing, because I had checked WHICH BINARY IS LIVE and never WHICH CODE PATH IS. 1161 // Logged on EVERY exit including the deadline kill, since a job that had to be killed is exactly 1162 // the one a reader needs to see. 124 mirrors the shell timeout convention. 1163 // Both rows carry the job's clock ROW NAME (2026-09-18): the organ column keeps argv[0], the name 1164 // rides as the last token, so a reader can tell WHICH row ran when many rows share one organ. 1165 if timedout == 1 { clk_actlog(path, 124, clk_slot(names, i)); return 1 } 1166 let raw: i64 = st[0]; let sig: i64 = raw & 0x7f; let code: i64 = (raw >> 8) & 0xff 1167 clk_actlog(path, code, clk_slot(names, i)) 1168 if sig != 0 { return 1 } 1169 if code != 127 { return 1 } 1170 return 0 1171} 1172 1173// load the registry into the arrays. returns #jobs (also written to np[0]). 1174func clk_load(path: *u8, names: *u8, organs: *u8, intervals: *i64, next_due: *i64, np: *i64) -> i64 { 1175 np[0] = 0 1176 let lenp: *i64 = sys_mmap(8) as *i64 1177 let data: *u8 = sys_read_file(path, lenp) 1178 if (data as i64) == 0 { return 0 } 1179 let dn: i64 = lenp[0]; var i: i64=0; var ls: i64=0; var n: i64=0 1180 while i < dn { 1181 if data[i] == (10 as u8) { 1182 if i > ls { if n < CLK_MAXJOBS { 1183 let line: *u8 = ((data as i64)+ls) as *u8; let ll: i64 = i-ls 1184 let d: *u8 = clk_slot(names, n); let e: *u8 = clk_slot(organs, n) 1185 var f: i64=0; var p: i64=0; var iv: i64=0; var nd: i64=0; var w: i64=0; var g: i64=0 // f: 0=name 1=interval 2=next_due 3=organ 1186 while p < ll { 1187 if line[p] == (9 as u8) { f = f + 1 } 1188 else { 1189 if f == 0 { if w < CLK_NAMEW-1 { d[w]=line[p]; w=w+1 } } 1190 else { if f == 3 { if g < CLK_NAMEW-1 { e[g]=line[p]; g=g+1 } } 1191 else { if line[p] >= (48 as u8) { if line[p] <= (57 as u8) { 1192 let dig: i64 = (line[p]-(48 as u8)) as i64 1193 if f == 1 { iv = iv*10 + dig } else { nd = nd*10 + dig } 1194 } } } } 1195 } 1196 p = p + 1 1197 } 1198 d[w] = 0 as u8; e[g] = 0 as u8 1199 intervals[n]=iv; next_due[n]=nd; n=n+1 1200 } } 1201 ls = i + 1 1202 } 1203 i = i + 1 1204 } 1205 np[0] = n 1206 return n 1207}