Every value write of the state was already a compare-and-set from the latest revision, retried on
conflict; the deletes were not. The sweep and the drop deleted a run's key whatever it held, so a
rerun's run replacing the key between the read and the delete was lost; the end of a run, the
ask-taken and a rerun deleted bound keys a newer run of the key may have armed. Each delete is now
at the revision just read, while the key holds the run it was judged from, judged again on a
conflict (the kind the runner's test 109 met in test-runs).
Builds: design 47 G §719 (readers added and removed by compare-and-set; the key deleted after composition-runs-kept), §720 (the bound keys)
The assigner said the same list of about 100 modules every two or three seconds, once per
pass (live on the stable node since 2026-11-17 11:02 CET, the second meeting); it now says
the list once each time it changes. Why their manifests cannot be read is the builder's
stream of build ends, which holds no end of a build made before 2026-11-11, said apart.
Builds: design 47 G G-23 (the assigner's long line, met a second time)
A module whose build can never be made left its machine not judged, so every step of the
candidate could not run; the case now carries none, the revision's words for each module it
gives no fingerprint, for the composer to blame that module alone.
Builds: design 47 G §721, §722; A-470j, A-470k
One test of the whole assigner: one run per candidate and machine read by every step, a
module's fault failing only its own steps, the bound of owed progress, the untaken ask, a
step-ended after the bound ignored, a rerun moving the reader that did not pass, a
superseded step leaving its run, a restart taking a run up by its ask, the drop verb and
the cap.
Builds: design 47 G §733 (test 108); A-470f
A bucket's listing misses a key that is being written at that moment (97 of 300 listings on
an embedded bus while one key was updated), so a run's end, its ask-taken or a moved reader
could miss the run being written. Every lookup of composition-runs now takes the keys the
state's own watch handed beside the listing.
Builds: design 47 G §719 (the assigner's own watch of its state), §720, §727; A-468, A-470d
The assigner fills it from the revision; the node-engine's composer reads it alone to tell a build that can
never be made, and judges the machine with that module held instead of leaving it not judged.
Builds: design 47 G §721, §722, A-470j, A-470k
A composition run must be seen and recoverable without a hand act on the bus: the seat
mesh-wants' verbs compositions and compositions-drop are answered by the process, the drop
recorded in the hand-act log first, its readers ended could not run naming it; the tools
assigner_compositions and assigner_compositions_drop call them.
Builds: design 47 G §728 (the assigner's two); A-469f, A-469u, A-469v
Steps asked again together start one run per machine they did not pass: a rerun keeps the
verdict of a machine it passed, joins a running run, reads one it did not read, or starts
a new run carrying every former reader, those that did not pass marked rerun with the
caller; at the new run's end each marked reader says its end again, naming the new ask,
and a reader that passed does not move.
Builds: design 47 G §727 (composition part); A-468g, A-469l, A-470d, A-470e
A step's end and its progress name per machine the composition run and its ask, so a
result is traced to the one ask that made it; a superseded or cancelled pipeline's
composition step leaves the runs it reads, the run going on for its other readers, and a
run left with no reader drops its ask and ends could not run.
Builds: design 47 G §723, §729 (mesh-wants' step-ended and step-progress shared); A-469n, A-469t, A-469z
A run whose build node stops saying progress, or whose ask no build node takes, would hold
its readers for ever: ask-taken arms bound.<run key> and each step-progress arms it again;
untaken.<run key> times the queue from the start, the readers said waiting for a build node
meanwhile; an expiry drops the ask and ends the run could not run naming the bound, and a
step-ended heard after it is ignored and journaled. Readers of a run left open at a restart
are taken up from the keys.
Builds: design 47 G §719 (take-up at start), §720 (the bound, when the clock ends, a run never taken, what the assigner hears); A-466q, A-466s, A-468, A-468b, A-469b, A-469s, A-469y, A-470
One module's fault fails only its own pipelines: a machine blaming module A fails A's step
and passes B's with a warning naming A; a machine that blames none fails every step on it;
a composer before blamed is judged as before, saying it names no culprits; a module the
search did not try is could not run for its own step. Each request carries blame_tries.
Builds: design 47 G §721 (an old composer, its bound), §722; A-466r, A-468a, A-469h, A-470b
The steps of one candidate share a machine's run, so a revision of 24 pipelines composes
each machine once: the step that starts a run asks the node's input, keeps one request,
chooses and records the composer program, and puts one run ask; every other step reads
the run's verdict from its key, its end written there from the build seat's step-ended.
Builds: design 47 G §718, §719, §720 (within its candidate only, could not run is never read); A-468, A-468a
A composition run is kept once per candidate and machine, so steps of one candidate share
it: the key is written create-only, a reader is the step's StepRun with its ask id, a run
that could not run is replaced, a run past composition-runs-kept goes at the next start,
and at the cap no run starts, naming the fill. The settings reach the process through
resource files.
Builds: design 47 G §718, §719, §720 (could not run is never read); A-468, A-468b, A-468c, A-469d, A-469w, A-469x, A-470
A step that read a composition run or a test run names it in shared, so the result is
traced to the one ask that made it; stands names what a reading step waits on; the
runs' read and drop are seat verbs with their class (seatgen regenerated, no hand edit).
Builds: design 47 G §728, §729, §731; A-468d, A-468e, A-469f, A-469q, A-469r, A-469t, A-469u, A-469v
The assigner bounds the composer's culprit search through the request, since the composer program has no settings; a
verdict always writes blamed, an empty list when it blames none, so an old composer's absent field stays tellable.
Builds: design 47 G §721, §729; A-469h, A-469i
The state inputs leaves the seat mesh-wants' definition (seatgen regenerated, no hand
edit). An ask's key holds {digest, at}; the sweep deletes an ask's key a day after its
time and an input with its object once no ask's key names it; keys of the old form
<time>.<digest> go by the time they carry, never earlier.
Builds: design 47 G A-465a, A-465b
The key of the state inputs is the digest alone (seat mesh-wants' key form <digest>);
each ask's name holds the digest, the input and its object end with the last ask that
names it or a day after its last write (at.<digest>); the contract files of the seat are
edited by hand (generator not found).
Builds: design 47 G A-465a
A step whose every machine is away never passes: it ends could not run, naming each machine and why. An away machine is still
not asked and still said as a warning when another machine is judged.
Builds: design 47 G A-467a, A-467
A machine the composer says `not judged` ends the step could not run, naming why (was a warning). A revision that names no seat
definition (A-234's stand-in) hands the composer every definition of the trunk seat list, read through mesh-catalog's
seat-definitions; a list or definition not read ends the step could not run, naming it (was an empty list, no error). The
integration bed makes the inputs object store, which a request carrying the seat list needs.
Builds: design 47 G A-467, A-234, A-252, A-261
Every step's end waited 2 s per retention (6 s) in the step log's Fetch when fewer than 20 entries lay past the aged pointer, so ends were taken one in 6 s serially.
Builds: design 47 G §330 (ageing of the runner's own data at each step's end)
At its start and on every source revision the indexer compares each repository's taken merge with the head the gate's source names; a repository behind is taken in as its merged message would have been, with the finding merge-untaken raised until it is.
Builds: design 47 G-22 A-463
A held have that carries a wait is a hold of the composition, not a failed backup; a backup ask whose have now carries a wait or is no longer held is withdrawn with its condition.
Builds: design 47 G A-453, ADR 0539 §16, §541
compose-candidate composes a whole revision in shadow on each node and
answered in about 34 s on g14 and shanks (2026-11-16 23:27), past the
SDK's 30 s call bound, so node-tools' composition step could not run. A
responder that is absent still answers at once. Quick fix; the thorough
one is a bound per verb from the seat's definition (after G).
main.981.node-tools passed build, lint, unit and integration; only its
composition step could not run (g14 and shanks composed in about 34 s,
past the assigner's 30 s, while shanks ran the pipeline's own tests).
The seat verb rerun asks one step again, but no tool reached it.
Builds: design 47 G §482
mesh-hand-acts 0.1.13 refuses a hand act with no cause; the pipeline
runner's record (cancel, rerun, accept, run, ask-again, review, start) and
the checker tools' requests drop gave none, so every one was refused and
nothing was done (runner_cancel, 2026-11-16 23:12: 96 of 96 refused). The
cause is the verb's own word, the one a second act for the same reason uses.
A batched take-in names only its newest revision's movers; the other modules it moved never got a pipeline.
Starter.TakenIn and the runner's first-half pickup now start one for each module of the branch's head with no pipeline at its fingerprint.
Builds: design 47 G §44, §54, §88, §476