sayHeld stays in G's judgeOverHeld, after the written-verdict branch, for a fresh trunk manifest
whose verdict the pass did not write. The test's merged manifest is written as the candidate's
own text ("2.0"): the join's manifest digest is of the raw bytes (meshcatalogclient
ManifestDigest), so "2" was another manifest and the test stopped before the case it adds.
Builds: design 25 bus-state §1 (merged from main)
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
Cause: the bed reports go1.27.1 for both nodes at its start, and the test then reports go1.27.2. The builder's new
durable on EVENTS hears the old reports first, so for a moment it agrees on go1.27.1 (said as toolchain-agreed
go1.27.1 by n1, seen in a run's EVENTS); a netcheck entry heard in that moment is built at its old fingerprint
("the old fingerprint, naming go1.27.1, was built with go1.27.2", integration_test.go:767, twice on 2026-11-12).
The builder acts as ADR 0465 says on what it heard; the bed told it a story the test does not mean. The test now
purges the bed's go reports before its own, and asserts that no toolchain-agreed of go1.27.1 is ever said.
Builds: design 47 G §492, §604, A-282
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
The floor wake (half.floor), the builder (half.b), ready and branches.Changed were set after reader.Start, so a
catalog view heard on the reader's goroutine read them while main wrote them: the race detector named
catalog.go:117 against main.go:501 in TestBuilderMakesADoneKeyWithTheDigests. They are now set before the start;
the builder cmd integration package then ran 21 times under -race with no race.
Builds: design 47 G §614, A-119
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
mesh/check being carried to 104909817b70 on 76db40eb6be1 | from 983c46e2576c
A pull request's candidate is judged before its merge and its verdict written, said nowhere. At the merge the
digest becomes the module's trunk manifest, but the merge's judging pass finds the same verdict, writes nothing
and so said nothing: the bus module never heard the new manifest, and the grants part, which follows the
verdicts on manifest-accepted, held the module as having no verdict and wrote no grant for its new fingerprint
(claude-licence-manager, three times on 2026-11-12). The merge's pass now says a fresh trunk manifest's accepted
verdict from the verdict held, with its revision.
Builds: design 25 bus-state §1, §12.2 (the indexer says manifest-accepted for every newly accepted manifest; the
grants part follows the verdicts on it)
Builds: design 25 bus-state L1 (hq ff30ac16, the lead's answer: a merged candidate's manifest is newly accepted).
c5a6871f created the key once and never wrote it again, so a step that kept moving was bounded from its first
movement (issue 1930). Arm now creates the key when absent; one ending no earlier is armed already; one ending
earlier is written again by an update against the revision read, carrying the bound's time to live; a waiter
whose update loses reads the key again and takes the newer bound as armed. The test moves a step twice past its
first bound, with two waiters each time, and reads one write per movement, its time to live, and no delete.
Builds: design 47 G A-282, A-115, §544
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
stepBounds.Arm deleted the key and created it again, so two builder instances waiting on one ask raced: one's
create met the other's and failed "wrong last sequence", and the ask's build failed (the builder bed's two
instances together). Arm now creates the key only if absent; one already there is armed, and nothing deletes
it to make it again.
Builds: design 47 G A-282, A-115, §544
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
An asker names the queue it asks as seat:<seat>.accept.<queue> (A-231); no module's serves lists that form, so
graph.Servers matched nothing and the indexer drew no served-verb edge from a build seat's holders to the asker.
Servers now reads the form through mesh-sdk's catalogue.QueueInvoke and answers every module serving a verb of the
seat; the kind-6 edge asks the module graph's Servers, the one rule, instead of a match of its own.
Builds: design 47 G §544, §605, A-231
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
gather asks the machines side by side, and the test's stand-in call wrote one map from each of them: a concurrent map
write, the panic once seen at compositioninputs_test.go:61 (the race detector shows it every run). The record is
guarded; the whole assigner's unit suite runs clean under the race detector.
Builds: design 47 G §354, A-243, A-252
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
The asks were made at a fixed time while the channel client checked their expiry against the real clock, so from
2026-11-12 every ask "expires before it is asked". The client now takes the same fixed clock.
Builds: design 47 G §568
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
The SDK's DeclareAskWorkers now declares each queue's two workers, main's and the named asks' (accept.<queue>.<asker>),
and the puller takes from both; the bed declared main's worker again filtering both forms, which overlaps the named
worker and the work queue refused ("filtered consumer not unique", 15 tests). The kind is a bed declaring by hand
what the SDK declares: the builder's bed was the only one whose own consumer overlaps the SDK's, and now declares
none. Behind it, the builder's seat server admits no caller of mesh-builds.reserved until the seat list holds
mesh-builds' definition (A-278); the bed's seat-definitions answers it from the contracts module's own file.
Builds: design 47 G §544, §600, A-115, A-278
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
A process's call gives the holder's own answer as its error (busretry.Answered); the entries kept by
fingerprint took that for a call that reached no holder and kept the entry asked, so §298's order edge
on the unread entry held the want for good and no pass ran again (the assigner's
TestWantsAreWrittenAndNoDifferenceIs, rule 7 "waits in place for the catalog entry").
Builds: design 47 G §614, §298, A-259, A-260
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
The bed's kept streams (running, machine-away, machine-back, version-delivered) took subjects of their own, so a test that
widened EVENTS met "subjects overlap". They now source EVENTS' messages of their subject, as the grants part makes them
(KeptEvent.StreamConfig), and EVENTS takes mesh.seat.*.event.>. The bed answers the seat mesh-catalog's seat-definitions
from mesh-sdk's seatlisttest copy and its verb entry from the catalog half, so the seat server admits a caller once the
list is read (A-278). The composition step test calls input on its subject naming the caller (A-179) and waits for the
seat list read and the sourced running messages. TestTheCompositionStepComposesEachMachineThroughTheBuildSeat passes;
TestWantsAreWrittenAndNoDifferenceIs still fails (rule 7: web on ace held on an unread catalog entry).
Builds: design 47 G §354, §610, A-179, A-278
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
A holder says a held call's end on its seat's gated-ended, kept last per subject for a day in SEAT_<seat>_gated-ended,
which the grants part makes from the definition seatdef.Resolve gives (A-264). The caller is granted hearing
gated-ended.<caller>.> under its own escaped account name, and an ordered consumer on that stream filtered to it, from its
invokes of the marked verb by name; a module or process that answers a marked verb is granted saying the event to any
caller. No process of mesh-delivery holds a marked verb, so no holder here takes NewGateChannel.
Builds: design 47 G §677, §700, A-264, A-273
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
The grants part grants a seat verb only to a module that names it in invokes (GR8), so a call missing from a
manifest is refused live. Listed every module's seat-verb calls in code (constant clients, seatlist.HolderOf,
operatorask, buildrun's drops and puts) against its invokes: the teller asks who holds operator-channel
(operatorask.Bucket) and the checker drops asks of the image build seat. Every asker of a build seat names the
queues it asks as seat:<seat>.accept.<queue> (A-231): the runner check and delivery, the assigner check, the
builder delivery, the checker check and recheck; the indexer had its entry. The validator judges that form
through mesh-sdk's catalogue.QueueInvoke against the kinds the seat accepts.
mesh-grants: a consumed seat view whose seat has no holder on record grants nothing instead of refusing the whole
module (TestGrants); TestCallersAndWhyRefused and TestSentStoresRead take §610's class defaults.
Builds: design 47 G §544, §601, §610, §616, A-231, A-179
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
The readers' invokes are read with mesh-sdk's catalogue.InvokedVerbs, so a long-form {verb, fields} entry (the
assigner's) is read rather than refused by the test's own decode.
Builds: design 47 G A-275, §558
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
For a module that requires finding-events and is bound, the membership lists the stream its binding names, the subject
mesh.bus.<module>.finding-events.{event}.<node> and the finding events, so the tool runner carrying a stdio module (the
enrolment) says an emitted finding event there; the test holds the enrolment's membership.
Builds: design 47 G A-280, A-274, §529
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
Each asks the seat mesh-broker's show for its finding-events entry (process.Binding, started without waiting), passes it
as the writer's Bindings and hears the seat's recreated and dropped; the watchman's one handler of the seats' events
hands those two to the binding, since a message has one taker. The teller, view and watchman manifests gain the invoke
seat:mesh-broker.show and consume mesh-broker.recreated and dropped. The forge bridge's and the watchman's whole tests
answer show and read the finding-written under the binding's subject.
Builds: design 47 G A-274, §529, §577, A-13
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
The template state takes max-mib (1 to 8192, as the old state section reads it). A state named approvals is refused
without key-ttl true, with ttl-seconds other than 172800 (approval.KeptFor, 48 hours), or with max-mib above 64; with
none declared its bucket takes the default 64 MiB.
Builds: design 47 G A-277, A-270, §677
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
G's grants already carry a state's max-mib onto its bucket (Bucket.MaxBytes, §469); main's second copy of that cap in
BucketStreamConfig and Composed is dropped for G's, and main's tests (the declared cap row, the held bucket raised)
are kept. The checker keeps G's version 0.1 (§368) with main's compositions max-mib 256.
Builds: design 47 G §469, §368; A-277 (since mesh-delivery #206 the grants part carries max-mib)
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
The finding pipeline-waits was raised at the wait's bound and nothing cleared
it. It now names the awaited runner fingerprint in one evidence entry,
`awaited <fp>`, written when it is raised (A-276), and the runner's hearing of
the node-engine seat's running clears exactly the findings whose entry names
the fingerprint said healthy (A-218), through its held writer, which holds what
its state findings held at its start: a finding raised before a restart is
cleared too, and one awaiting another fingerprint stands. The summary is never
read for it. A clear that fails has the running heard again.
Builds: design 47 G lead answers A-218, A-276 (join 11 area 3)
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
The builder's published package reached the bus library through buildstep,
whose KeptBuildsOf took mesh-sdk's buscalls.ObjectInfo, and buscalls now
imports nats.go for its seat chunks. buildstep's layouts, keys and codecs
are read by modules that never touch the bus (its own guard), so the kept
build of one object is judged from the object's own fields (KeptBuildOf,
SortKept) and the build module lists its store itself. The guard now also
refuses a bus library reached through an import, the kind that let this in.
Builds: design 47 G (join 11 area 4: builds TestThePackageReachesNoBus, buildstep's bus-free guard)
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
The five §298 order-wait tests of b2l (A-245, A-258, A-259, A-260) held no
wait and said none after the merge with b2f: b2l's fixture put ace's presence
as online (presenceValue(true)), and the merge rewrote it as b2f's
heardPresence("ace", true, ...), whose flag says away. With every node away
no module had a canary, so no want moved and no order edge was held (an order
wait starts only on an online node, §268). The fixture now hears ace back;
the order waits are built as those answers have them, unchanged.
Builds: design 47 G lead answers A-245, A-258, A-259, A-260 (join 11 area 2)
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
AnswerSeat checks a call's caller against the seat's definition its caller
reads from the seat list it follows (A-186), refusing every caller while the
list is unread. The assigner (mesh-wants input), the builder (mesh-builds
reserved) and the indexer (mesh-catalog input) hand it the definition their
seat list holds once heard; the pipeline runner, which followed none, follows
one over its calls connection and hub, read again once its events are live,
and its manifest invokes seat-definitions and consumes the list's two events.
The runner's and the indexer's integration beds answer the seat list, and the
indexer's input test calls with a caller the bus proved (A-179).
Builds: design 47 G lead answers A-179, A-186 (join 11 area 6)
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
The assigner keeps each reads-changed event it hears beside its statement (enginereads.Keeper.HeardChanged),
reads a node as the event says while its statement is behind it, and asks again until the statement is as new.
Builds: issue 1975 note 4 (the lead's handover; no design line)
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
The case main's validator release added names `version` "1", which G's
reader refuses (MAJOR.MINOR, design 47 G §368); it names 1.0.
Builds: design 47 G §368, §508
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
The join's importers name the delivery repository's moved contracts and sdk
(MV1), node-tools' tests use mesh-sdk's gate, the duplicate node-engine claim
rule of the reader release gives way to G's, and the assigner names its first
watched branch, never a constant. The seat holders' AnswerSeat callers are
area 6's.
Builds: design 47 G join 11 (merge of fix round 11)
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
The validation run's answer is read from the step's end through steprun's
ReadOutput, whole or from the build seat's output object, and the indexer
invokes seat:node-build-go.output for it (A-254); the patch B5g6 left for
the join, applied on the merged heads.
Builds: design 47 G §712, §713; lead answers A-254, A-268
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
Its head stands where a merged message's merge stands: forgeKind, pending refs, the newest load and the revision's head.
Builds: design 47 G §509, §714, A-269
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
Reverts the Go-bundle package of 53caa3ba; the run's program is the directory ./cmd/mesh-judge under the go.mod the build reads.
Builds: design 47 G §712, §676, §491, A-268
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
jschoubben