mesh/merge-gate pass: builds mesh-host → ace, g14, novox, shanks; no bus step; every machine composes with the change as it did without (4 of 4 compose)
The review of #79 found that pacing alone brought back issue 361's
fault. The control-node's own I/O pressure sits near the 10 % bound with
no walk running, so a walk that paused 30 s for every 250 ms it read
would read 0.8 % of the time: a 16-minute walk would take some 32 hours,
outlive the age a search judges for, and be begun again at an engine
restart, leaving the machine unjudged.
A run of pauses is now at most nine times as long as the walk read since
the one before, so the walk keeps at least a tenth of the time whatever
the pressure, and takes at most ten times as long as unpaced: about
2.7 hours for the 16-minute walk, inside the 12 hours the controller
waits quietly. A test ties the bound to rootsearch.Quiet.
The pressure is the share of the time since the last look in which some
task waited on I/O, from PSI's cumulative counter, not its ten-second
average, which the kernel moves only every two seconds. A pause is
500 ms, then the pressure is looked at again.
The engine's walker is built in one place, engineWalker, and a test
fails if it is not paced by /proc/pressure/io. What the walk cannot read
is said through the search's own channel (the cache's Say), as the rest
of the search is, with what it means for the disk.
mesh/merge-gate pass: builds mesh-host → ace, g14, novox, shanks; no bus step; every machine composes with the change as it did without (4 of 4 compose)
The walk for setuid programs is put in the idle I/O class so it yields
to everything else, but only a scheduler that schedules by class (BFQ)
honours that. The control-node's disks run mq-deadline, which does not,
so the walk read at full priority. On 2026-10-11 the walk started at
01:27, right after the engine applied a new controller, and ran sixteen
minutes; it kept two spinning disks saturated, the bus whose store
shares them stalled for half a minute twice, and the controller lost
its lease and restarted each time.
The walk now also paces itself by PSI, which reads the same whatever
the scheduler: before it reads a directory, while some task has waited
on I/O for more than 10 % of the last ten seconds, it pauses two
seconds at a time. It looks at most every 250 ms, pauses at most 30 s
in a row and then reads one more directory, so a machine that is never
quiet is walked slowly but still judged. Its own reads count in the
pressure, so on an otherwise idle machine it holds the disk below the
bound. Pressure it cannot read pauses nothing, and is said once.
A paced walk finds what an unpaced one finds; a walk ended while it
pauses is an incomplete walk, never one that found nothing.
mesh/merge-gate pass: builds mesh-host → ace, g14, novox, shanks; no bus step; every machine composes with the change as it did without (4 of 4 compose)
The two finds on the control-node sit in the operator's home of mode 0700, which the agent account cannot
search. The search keeps every find; each judgement reads the modes on the way again, since an account can
open its own home without root. A find in a home is said by count, its path in the journal alone. Also
from review: a checkpoint racing SearchNow forgets the walk it just kept, and a walk kept by an older engine
says its counts as lower bounds.
mesh/merge-gate pass: builds mesh-host → ace, g14, novox, shanks; no bus step; every machine composes with the change as it did without (4 of 4 compose)
The review found a cursor of a name that is not UTF-8 came back changed
from JSON, so a resumed walk skipped what it stood before; an ask to
search again was lost to a restart, which carried the older walk on; a
fresh search judged a filesystem it never walked; a mount point root
cannot read failed or hung the walk; a found file gone before the package
manager was asked failed every resume; and skipped or kept outcomes
started a search on every reconcile, back to back with a long walk.
mesh/merge-gate pass: builds mesh-host → ace, g14, novox, shanks; no bus step; every machine composes with the change as it did without (4 of 4 compose)
The control-node's walk takes far longer than the engine runs between
updates, so a search that starts over at every restart never completes.
The walk is the engine's own, in lexical order, so its progress is one
cursor per filesystem, kept every minute beside the engine's state and
carried on after a restart while younger than rootsearch.FreshFor. It runs
on a thread of its own at nice 19 and the idle I/O class.
A machine with millions of files never finished the walk within fifteen
minutes, so its agent account was never judged and the router refused to
link. Only root can make a setuid-root file, so a complete search stays
sound until root acts: it now runs at idle priority with no bound, is kept
beside the engine's state, judges for rootsearch.FreshFor from its start,
runs again daily and after an apply that moves accounts, packages, actions
or sudo rules, and an incomplete or stale search is never free.
mesh/merge-gate pass: builds mesh-host → ace, g14, novox, shanks; no bus step; every machine composes with the change as it did without (4 of 4 compose)
SetuidCache compared the caller's now with a time read from the wall clock when the search finished. When the caller's clock ran behind the wall clock, a search was never due again: TestTheSetuidSearchRunsInTheBackgroundAndIsKeptForItsInterval, whose clock is a fixed date, has failed on main and every branch since 2026-10-09 18:00 UTC. A search's time is now the now it was started at; a result reads at most one search older than it is, which errs towards searching again.
mesh/merge-gate pass: builds mesh-host → ace, g14, novox, shanks; no bus step; every machine composes with the change as it did without (4 of 4 compose)
The search kept a two-minute limit of its own inside its 15-minute bound, so
a long walk always ended incomplete; pacman was asked once per path; and find
exiting with an error but some output was read as a complete walk. The bound
and the not-judged reasons move to the public rootsearch package so the
controller reads the same values (the review of 2026-10-09).
mesh/merge-gate pass: builds mesh-host → ace, g14, novox, shanks; no bus step; every machine composes with the change as it did without (4 of 4 compose)
A search over a large disk ran inside the look, holding the judge's lock and
stalling the health statement; it now runs apart, serving its last result,
backing off after a failure, and saying not judged until it has one. And a
rule naming a user, or a comment, no longer hides another rule's
unconditional yes (hq ADR 0266).
A hard link swapped in for ~/.claude would have had root chown another
account's file; fstat on the descriptor now refuses a second link, a fifo or
an unexpected kind before anything is changed (hq ADR 0266, the re-review).
The judge also finds polkit rules for every account, a runtime's API on TCP,
setgid-to-root programs whoever owns them, setuid programs on every suid
filesystem, and unprotected links; the rest is listed as not judged.
An account could replace ~/.claude with a link to /etc and have the
node-engine chown, chmod or write through it on the next apply. Below a
person's or an agent's home every component is now opened without
following a link, and a link refuses the resource in words.
The root judge also reads doas and polkit rules, the container runtimes'
sockets with their ACLs, ACLs on the secrets, and setuid-root programs no
package owns, in the C locale; Judged and NotJudged write down exactly
what it covers (hq ADR 0266 review).
jschoubben
jochen