Design 47 H: split the lead answers; H2's (H-A-373 on) in H2-lead-answers.md

This commit is contained in:
2026-11-22 00:15:30 +01:00
parent 6d90185f49
commit dfbca1458d
2 changed files with 187 additions and 180 deletions
@@ -2315,183 +2315,4 @@ Where two preparers disagreed, the chosen one is named.
with an empty version, so the builder says "package-published not said … a by-digest publish" and the bundle is
not marked against its reservation. Fix in the Node build module: the publish carries the version the ask
reserved. Test: a bundle's publish names the ask's version. P3, P-failure-loud.
- **H-A-373 (every build module's publish names the ask's version, 23:42)** The developer of H-A-372 found the Go build
module publishes every bundle and archive with an empty version too (its publish of an archive). The second of a
kind: a build module's publish that names no version marks no reservation. Fix the kind: every build module's
publish (the Go build module's bundles and archives, the Node build module's bundles, the image build module's
images where it publishes) carries the version the ask reserved, built in the same change as H-A-371 and H-A-372; the delivery SDK's publish refuses an empty version
with a Go bundle or package, loudly, so no build module can publish by digest alone again: that part is H2's
mesh-delivery developer's, after its Left list. fix-a-bug-met-twice, P-failure-loud.
- **H-A-374 (a directory's declared mode is applied to a directory that exists, 23:42)** Pipeline main.175 failed at
23:24 because the Node build module's integration tests, asked of the Go build seat on ace, could not see the Node
toolchain: ace's toolchain cache directory is 0700 on disk while the manifest's directory entry says 0711 (ADR 0533)
and the other build nodes have 0711; the node-engine's composition on ace says "unchanged, equal to the composition
the node runs". So the node-engine applies a directory entry's mode when it makes the directory and never when the
directory exists with another mode, and its judgement reads what it wrote, not the disk. Fix in the node-engine
(mesh-host, H2's E lines): applying a directory entry sets its owner and mode when they differ on disk, and the
composition's diff names a directory whose mode on disk differs. Test: a directory that exists with another mode is
set to the declared one at apply. Recovery tonight: the mode set through the files seat's chmod verb on ace with
this answer as its reason; the engine's next apply must leave it. P-mesh-owns-the-truth, P-failure-loud.
- **H-A-375 (a pipeline step a node cannot run moves to a node that can, 23:42)** The same failure showed a gap: the
Go build module answered the integration step's digest "left to a node with node-build-node: this node lacks it",
the step ended could not run, and the runner counted that as failed and ended the pipeline; nothing placed the step
on g14 or shanks, which have the need. ADR 0552 §2 and §4 decide this for check asks (a holder lacking a need
declines; the asker places the unmet packages again on a node with the need) and no design line carried it to the
pipeline's step asks. Carry it: the runner's step ask names the packages' needs (read from the module's needs
marker, as the Go build module does for its digest); a build seat holder lacking a need declines the ask instead of
ending it could not run; the runner places a declined step again on a node that has the need, once per node at
most, and a step no node can run ends could not run naming the need and the nodes asked. Built by H2's
mesh-delivery developer (the runner's placing and the delivery SDK's step run) after its Left list, in H2's one
review; it is not a small fix. Until then a could-not-run step is rerun by hand (runner_rerun). P-NATS-native,
P-failure-loud, ADR 0552.
- **H-A-376 (the runner's list answers a short form, 23:53)** The operator asked where to follow the pipelines; the
runner's list of main answers 2.3 MB, above what the bus carries, and show of one pipeline 127 KB, so the state can
only be read narrowed by module, mover and limit, and stands has no description. One question, one tool: list
answers by default a short row per pipeline (key, call, started, ended, outcome, the movers' count, the first failed
or could-not-run step with its module and why), the newest 20, and the whole pipelines only with `full: true` and a
limit; show answers the steps of one mover when given, else the pipeline's rows per mover (module, outcome, the
failed step) and never every step of every mover; stands, kept and why-carried get the description their seat
contract names. Built by H2's mesh-delivery developer after its Left list. tools-are-how-a-module-is-spoken-to,
one-question-one-tool, P3.
- **H-A-377 (drain-bound is a duration, like every time setting, 23:57)** H-A-358 made the setting drain-bound a whole
number of seconds in the node-engine, and H-A-370 says the node-engine composes a duration setting as its text ("10m")
for the one typed reader: the two meet on the tool runner, where stdio.SettingDuration refuses "30". H-A-370 holds:
a setting that is a time is a duration text everywhere ("30s"), declared, composed and read so; the node-engine
parses it as a Go duration and derives whole seconds only where it renders one for systemd (TimeoutStopSec takes
"30s" as well, so no derivation is needed there either), and its refusal of a non-integer ends. H-A-358's "whole
seconds" is the rendering, never the setting's form. The runner reads drain-bound with stdio.SettingDuration, 30 s
default; no seconds reader is added to mesh-sdk. The node-engine's developer changes drainBoundOf (E3's
counterpart), the tool runner's developer reads it. P9 (one reader, one form), H-A-370.
- **H-A-378 (S8 changes no mesh-sdk code, 00:02)** §783's finding events (the events template, 8 MiB drop-oldest,
findings-max-age) are each module's manifest entry; the writer already says its events under the finding-events
binding when bound. S8 in mesh-sdk is done by what exists: no refusal in the writer, no exported constants (a
manifest is text the SDK's constants cannot reach). The one place that holds the numbers is §814's manifest
validator in mesh-catalog (the finding-events binding's cap 8 MiB, drop-oldest, the setting findings-max-age with
its 7-day default), and the catalogue's module.json rows carry them. The mesh-catalog developer builds the validator
rule with the rows; the mesh-sdk developer closes S8 as built. A rule states how it is checked (hq), P9.
- **H-A-379 (an install's end is a step-ended naming the step install, 00:00)** H-A-367 (4) and §800 give the Node
build seat an install whose end carries the archive object's name; the build-step part's step enum knows lint, unit
and integration, and the delivery SDK refuses a step-ended naming another. The step enum and buildstep.PipelineSteps
gain `install`: every pipeline ask ends in a step-ended naming its step, one way, and the archive object's name rides
the install's end as §800 says. The contract part (mesh-contracts, regenerated) and the delivery SDK's check change
together in the group; the Node build module's install answers it. P3.
- **H-A-380 (the prune helper for a process, the keep's home, the tools' names, 00:08)** §812, S11. (1) A process's
state write over the bus's key-value put answers a revision and no fill (the bundle side gets the Fill the runtime's
put answer carries, DMC-14). The helper for a process reads the entry's figures once when it binds the state and once
after each prune, and between those keeps a running sum of the bytes each write puts (an upper bound: the bucket's
history bounds what an overwrite keeps); the sum passing `prune-mark` of the cap is the judgement. No read per
write, no poll, no timer: H-A-227's "read it once", H-A-344 (7). (2) A keep lives in the pruned state itself, under
the reserved prefix `.keep/<key>` with its until, never pruned and never counted as an entry by the readers; the
prune reads the keeps once per prune and skips a key kept until a time still to come; the hand act is written in the
hand-act log as §812 says (the audit trail, not the truth). Oldest first is by each key's current revision, lowest
first. No new store per module (one store per thing, P3). (3) The tools are named as the generic tools are:
`<module>_state_prune_now` and `<module>_state_keep` (stateobjects.Tools' form); §812's bare names are the tool
kind. P3, P9, P-NATS-native.
- **H-A-381 (a repeat of a taken ask, and the taken key's bound, 00:12)** §775, RG1 and C2. (1) A holder that is asked
again with an id it took follows H-A-364 (b): while the first act still runs the repeat joins it and gets the same
answer; after it ended the repeat answers the act's state as it stands, read afresh (the registry's publish answers
the transfer's parts held; a git verb answers the pull request as it is: merged, with its merge commit), which is
what the act answers when done; the act is never performed twice and no answer is kept in `taken`. A repeat for
which the state says the act never happened (the transfer is gone) is a refusal naming it. (2) The holder never
receives the caller's bound, so the taken key's TTL is a setting of each holder, `ask-taken-bound`, default 10
minutes (the largest caller's bound, the runner's ask-unanswered-bound), declared in the registry's and the forge
module's manifests; the node-engine keeps compose-ask-bound as H-A-364 (a) decided. §775's "the ask's bound" reads
so. P3, ADR 0668 §2, principle 11.
- **H-A-382 (a page by cursor: seen-again and each kept message, 00:16)** §779, T1. (1) The runner keeps nothing per
read and a bucket keeps no trace of a key's removed older message, so seen-again marks every message written after
the read's first page (the cursor carries that page's stream sequence), a key first created after the read began
included: no state per read in the runner's memory (the mesh owns the truth, P-NATS-native), and a reader replacing a
value it does not hold loses nothing. (2) A page hands each kept message in sequence order, a key's older kept
revisions included, each later revision of a key already handed marked seen-again by (1); only the first page may
ask last-per-key, and it does not: one way for every page (P3), as §779's acceptance says ("delivers each kept
message once"). P3, P-NATS-native, H-A-227.
- **H-A-383 (the node-engine's taken state names its cap; the baseline gains the word, 00:20)** E2. §820 gives every
holder's `taken` 8 MiB, and every kept store names its cap (§812, ADR 0595); the node-engine is no exception. Its
manifest declares taken {history 1, key-ttl, max-mib 8}. The word `state.max-mib` is one the engine reads already
(it writes every module's bucket cap from it), so mesh-sdk's baseline manifest reads (ADR 0614 §4) were incomplete:
the baseline gains `word:state.max-mib`, in the same group, and packaging/operator_keys_test.go follows the baseline.
The running engine on each node holds the new engine's build for a yes once, at this delivery, because its baseline
lacks the word: that hold is said to the operator before the delivery starts and answered then; it is the delivery
system doing what ADR 0614 asks, not a reason to leave the store uncapped (design-it-as-it-should-be).
P-failure-loud, §812, ADR 0614 §4.
- **H-A-384 (three node-engine lines the developer thought wrong, 00:24)** (1) §813 (4), H-A-357: the nft comment is a
label, the composition is the truth. Past 128 bytes the module list is cut with a count, `derived a,b,+4 tcp/443
scope any`, the whole list standing in the composition's filtering resource; the holder's rules reading (mesh-catalog
P1) reads that form; the compose never refuses for a label's length (a refusal would take the whole table down for a
shared port: not proportionate; P-failure-loud is served by the composition, not the label). (2) §813 (1), H-A-355:
the applied run is kept in the engine's have, the record of what it applied that it already writes to the bus (the
composition carries no run, as H-A-355 says); a restarted engine reads its have and never applies an older run again.
Memory is never the truth (P-mesh-owns-the-truth). (3) §813 (12), H-A-358: drain-bound raises the declaration's
form (the process's form rises as DMC-16 raised the build-step part's), so the delivery plan derives the ordering
edge node-engine (the reader of the new form) before the composer that writes it; the older engine's strict decode
stays strict. The edge is named in the milestone plan with this answer (design-it-as-it-should-be). P3,
P-mesh-owns-the-truth.
- **H-A-379 (addendum, 00:24)** The install ask carries no artifact or fingerprint, so the delivery SDK's check takes an
install's end without them, naming module, commit and the archive (mesh-delivery 1767d7b4).
- **H-A-385 (the gate's ledger entries: what each keeps, 00:34)** §809, GT1. (1) The head entry `<branch>.<source>`
keeps every field about the record as a whole: group, trace, acts, away, recovery, call, the pull requests' heads,
with its outcome, the movers' count and §810's times; nothing per module. (2) A mover entry
`<branch>.<source>.<module>` keeps that module's step outcomes per node (taken-in, built, checked, composed, wanted,
running-healthy, held, failed, away, each with why and event id) beside its pipeline, version, machines and running
times: §809 forbids an entry naming two modules, not a module's own detail, and delivery_status and show read them
there; the 64 KiB cap holds per entry. (3) The catalog half, when it moves, is a mover with an entry of its own
(`<branch>.<source>.mesh-catalog`, no node) and counts in the head's movers. (4) The month list (type held) stays
the way the ledger is listed; nothing changes it. One way per thing, P3; ADR 0595 §3.
- **H-A-386 (the keep's prefix is `_keep/`; the runtime's put answer carries the fill, 00:14)** Corrects H-A-380 (2):
the bus refuses a key-value key beginning with '.', so the keep lives under the reserved prefix `_keep/<key>`, the
underscore form the tool runner already uses for `_meta` (H-A-351); everything else of H-A-380 stands. DMC-14, which
H-A-350 answered and no build map carries: the tool runner's put answer (mesh/state.put) carries the entry's fill as
`{"revision": n, "fill": {"count", "bytes", "max_bytes"}}`, buscalls.Fill's form, which mesh-sdk's bundle-side prune
helper reads; built by the mesh-tools developer in H2 beside T1, tested in the runtime's state tests. P3, H-A-227.
- **H-A-387 (the artifact registry takes a publish whose reservation names no digest yet, 00:20)** Since H-A-373 the Go
build module publishes an archive with its reserved version during the build step, when the builder's reservation
still holds an empty digest for every artifact (the builder records digests only at the build's end); the
distribution module's holder (the seat oci-artifact-registry) accepts only an equal digest and says "for another
digest" of an empty one, so every archive-only publish since 00:05 was refused (twelve modules on three nodes, nftables
among them). A-490j decided this on 19 Nov for every holder: a publish is taken when the reserved artifact names that
digest or names none yet; commit ac2cd374 kept the strict check in the distribution holder against it. Fix in the
distribution module (mesh-catalog): holder.Reserved.Recorded takes an empty reserved digest as A-490j says, and its
refusal names the case it met ("reserved for digest X, not Y" or "no reservation of version V"), never "another
digest" for none. Test: a publish against a reservation with no digest yet is taken and marked; one against another
digest is refused naming both. The pack is reproducible (the same archive bytes at 0.1.0, 0.1.1 and the refused
publish). Recovery: rerun the build steps once the distribution module runs on novox. A-490j, P-failure-loud,
fix-a-bug-met-twice (the same symptom met 19 Nov).
- **H-A-388 (the have carries the applied run; the declaration's version is its form, 00:26)** E4b, E7b. (1) The
node-engine's have value gains `run` (the composition run's sequence the module was applied from) in
seats/node-engine/seat.json, generated, its form raised by one as H-A-352 (2) has it; the node-engine's developer
makes that K row in the mesh-contracts worktree (one commit citing this answer) since no other developer holds it
tonight, and tells the lead both heads. (2) The declaration carries no form beside its Version, so Version is its
form (one field per thing, P3): mesh-sdk's declaration gains the process's drain-bound under Version raised by one,
the composer (the assigner, mesh-delivery) writes that Version, the node-engine reads it and refuses a higher one as
it does today. The edge the delivery plan derives: the node-engine's manifest states the declaration Version it reads
(`declaration-reads`, beside its manifest reads of ADR 0614) and the composer's manifest the Version it writes
(`declaration-writes`); the plan orders the writer of a Version after every node's engine reads it. Built: the field
and the two manifest words in mesh-sdk (its developer), the engine's word and reader (mesh-host), the composer's word
and writer and the plan's edge (mesh-delivery, with A1-A6). Until the plan reads the edge, the milestone plan names
the order by hand: node-engine before the assigner. design-it-as-it-should-be, P3, ADR 0614.
- **H-A-389 (a validate step judging a candidate that changes a seat's form, 00:26)** The mesh-catalog developer found
the trunk validator refuses H2's delivery seat definitions (form, max-age: fields its format does not know), so a
pipeline's validate step, asked of the running assigner, cannot judge a candidate that raises a seat's form. Today
validate is an optional step (main.175 shows it so), so the H2 group's pipeline is not failed by it; the gap stands:
a candidate that changes mesh-delivery-sdk or a seat's form is judged by a validator built from the candidate's own
heads (ADR 0573 §2's rule for the check, carried to the validate step: the runner builds the validator tool from the
candidate like any artifact and asks it). Queued for H2b in mesh-delivery; the H2 group's validate verdicts are
read by hand once after its merge. design-it-as-it-should-be, ADR 0573 §2.
- **H-A-390 (§799's shared parts are the delivery SDK's developer's, 00:26)** BM1: the shared parts in
mesh-delivery-sdk's steprun (again-delays parsed into Puller.Delay, the third run's "fetch:"/"publish:" why,
HolderSilentOr's assigner call "with no seat call to learn why") are built by the mesh-delivery developer with its B
lines; the mesh-catalog developer builds the three build modules' settings and their use of the conformance kit
against the relayed head. One owner per file (P3).
- **H-A-391 (the gate's mover entries are found by an index on the record's key, 00:30)** §809, GT1. The ledger type
`mover` declares an index on its record's key in the gate's manifest `requires` (records' `indexes`), and
gate_record's join is one `mesh/ledger.query` on that index with the key `<branch>.<source>.>`; the head entry stays
per record (its count of movers and nothing per module), and nothing reads the revision's graph again. A mover's
instance path is `mover/<token of branch.source.module>` beside `record/<token of branch.source>`. P-NATS-native, P3.
- **H-A-392 (what H-A-374 changes in the node-engine, 00:30)** (a) Issue 339's rule stands: a directory the mesh never
made, found with another owner or mode, is used as found and said so; H-A-374 holds for a directory the engine's
record says the mesh made. (b) Two things, both read-only on the judgement: the verb composition with diff reads a
record-held directory's owner and mode on disk and names one that differs as drift; and the engine converges the
owner and mode of every record-held directory entry at its start and whenever it applies that module, one journal
line per directory set right, so a drift lasts at most until the engine's next start. An unchanged composition is
still not applied (idempotence keeps). Test: a record-held directory with another mode is set right at the engine's
start; one the mesh never made is left and said. P-mesh-owns-the-truth, issue 339.
- **H-A-373 and later** are in H2-lead-answers.md (this file was split at 00:40 CET on 22 Nov; it ends at H-A-372).
@@ -0,0 +1,186 @@
# Design 47, milestone H2: lead answers
The lead's one-line answers to the questions H2's develop and review raise (implement-the-design), from H-A-373 on; H-A-1 to H-A-372 are in H-lead-answers.md. Each names the principle that settles it.
## From H2's develop (2026-11-21 from 23:42 CET)
- **H-A-373 (every build module's publish names the ask's version, 23:42)** The developer of H-A-372 found the Go build
module publishes every bundle and archive with an empty version too (its publish of an archive). The second of a
kind: a build module's publish that names no version marks no reservation. Fix the kind: every build module's
publish (the Go build module's bundles and archives, the Node build module's bundles, the image build module's
images where it publishes) carries the version the ask reserved, built in the same change as H-A-371 and H-A-372; the delivery SDK's publish refuses an empty version
with a Go bundle or package, loudly, so no build module can publish by digest alone again: that part is H2's
mesh-delivery developer's, after its Left list. fix-a-bug-met-twice, P-failure-loud.
- **H-A-374 (a directory's declared mode is applied to a directory that exists, 23:42)** Pipeline main.175 failed at
23:24 because the Node build module's integration tests, asked of the Go build seat on ace, could not see the Node
toolchain: ace's toolchain cache directory is 0700 on disk while the manifest's directory entry says 0711 (ADR 0533)
and the other build nodes have 0711; the node-engine's composition on ace says "unchanged, equal to the composition
the node runs". So the node-engine applies a directory entry's mode when it makes the directory and never when the
directory exists with another mode, and its judgement reads what it wrote, not the disk. Fix in the node-engine
(mesh-host, H2's E lines): applying a directory entry sets its owner and mode when they differ on disk, and the
composition's diff names a directory whose mode on disk differs. Test: a directory that exists with another mode is
set to the declared one at apply. Recovery tonight: the mode set through the files seat's chmod verb on ace with
this answer as its reason; the engine's next apply must leave it. P-mesh-owns-the-truth, P-failure-loud.
- **H-A-375 (a pipeline step a node cannot run moves to a node that can, 23:42)** The same failure showed a gap: the
Go build module answered the integration step's digest "left to a node with node-build-node: this node lacks it",
the step ended could not run, and the runner counted that as failed and ended the pipeline; nothing placed the step
on g14 or shanks, which have the need. ADR 0552 §2 and §4 decide this for check asks (a holder lacking a need
declines; the asker places the unmet packages again on a node with the need) and no design line carried it to the
pipeline's step asks. Carry it: the runner's step ask names the packages' needs (read from the module's needs
marker, as the Go build module does for its digest); a build seat holder lacking a need declines the ask instead of
ending it could not run; the runner places a declined step again on a node that has the need, once per node at
most, and a step no node can run ends could not run naming the need and the nodes asked. Built by H2's
mesh-delivery developer (the runner's placing and the delivery SDK's step run) after its Left list, in H2's one
review; it is not a small fix. Until then a could-not-run step is rerun by hand (runner_rerun). P-NATS-native,
P-failure-loud, ADR 0552.
- **H-A-376 (the runner's list answers a short form, 23:53)** The operator asked where to follow the pipelines; the
runner's list of main answers 2.3 MB, above what the bus carries, and show of one pipeline 127 KB, so the state can
only be read narrowed by module, mover and limit, and stands has no description. One question, one tool: list
answers by default a short row per pipeline (key, call, started, ended, outcome, the movers' count, the first failed
or could-not-run step with its module and why), the newest 20, and the whole pipelines only with `full: true` and a
limit; show answers the steps of one mover when given, else the pipeline's rows per mover (module, outcome, the
failed step) and never every step of every mover; stands, kept and why-carried get the description their seat
contract names. Built by H2's mesh-delivery developer after its Left list. tools-are-how-a-module-is-spoken-to,
one-question-one-tool, P3.
- **H-A-377 (drain-bound is a duration, like every time setting, 23:57)** H-A-358 made the setting drain-bound a whole
number of seconds in the node-engine, and H-A-370 says the node-engine composes a duration setting as its text ("10m")
for the one typed reader: the two meet on the tool runner, where stdio.SettingDuration refuses "30". H-A-370 holds:
a setting that is a time is a duration text everywhere ("30s"), declared, composed and read so; the node-engine
parses it as a Go duration and derives whole seconds only where it renders one for systemd (TimeoutStopSec takes
"30s" as well, so no derivation is needed there either), and its refusal of a non-integer ends. H-A-358's "whole
seconds" is the rendering, never the setting's form. The runner reads drain-bound with stdio.SettingDuration, 30 s
default; no seconds reader is added to mesh-sdk. The node-engine's developer changes drainBoundOf (E3's
counterpart), the tool runner's developer reads it. P9 (one reader, one form), H-A-370.
- **H-A-378 (S8 changes no mesh-sdk code, 00:02)** §783's finding events (the events template, 8 MiB drop-oldest,
findings-max-age) are each module's manifest entry; the writer already says its events under the finding-events
binding when bound. S8 in mesh-sdk is done by what exists: no refusal in the writer, no exported constants (a
manifest is text the SDK's constants cannot reach). The one place that holds the numbers is §814's manifest
validator in mesh-catalog (the finding-events binding's cap 8 MiB, drop-oldest, the setting findings-max-age with
its 7-day default), and the catalogue's module.json rows carry them. The mesh-catalog developer builds the validator
rule with the rows; the mesh-sdk developer closes S8 as built. A rule states how it is checked (hq), P9.
- **H-A-379 (an install's end is a step-ended naming the step install, 00:00)** H-A-367 (4) and §800 give the Node
build seat an install whose end carries the archive object's name; the build-step part's step enum knows lint, unit
and integration, and the delivery SDK refuses a step-ended naming another. The step enum and buildstep.PipelineSteps
gain `install`: every pipeline ask ends in a step-ended naming its step, one way, and the archive object's name rides
the install's end as §800 says. The contract part (mesh-contracts, regenerated) and the delivery SDK's check change
together in the group; the Node build module's install answers it. P3.
- **H-A-380 (the prune helper for a process, the keep's home, the tools' names, 00:08)** §812, S11. (1) A process's
state write over the bus's key-value put answers a revision and no fill (the bundle side gets the Fill the runtime's
put answer carries, DMC-14). The helper for a process reads the entry's figures once when it binds the state and once
after each prune, and between those keeps a running sum of the bytes each write puts (an upper bound: the bucket's
history bounds what an overwrite keeps); the sum passing `prune-mark` of the cap is the judgement. No read per
write, no poll, no timer: H-A-227's "read it once", H-A-344 (7). (2) A keep lives in the pruned state itself, under
the reserved prefix `.keep/<key>` with its until, never pruned and never counted as an entry by the readers; the
prune reads the keeps once per prune and skips a key kept until a time still to come; the hand act is written in the
hand-act log as §812 says (the audit trail, not the truth). Oldest first is by each key's current revision, lowest
first. No new store per module (one store per thing, P3). (3) The tools are named as the generic tools are:
`<module>_state_prune_now` and `<module>_state_keep` (stateobjects.Tools' form); §812's bare names are the tool
kind. P3, P9, P-NATS-native.
- **H-A-381 (a repeat of a taken ask, and the taken key's bound, 00:12)** §775, RG1 and C2. (1) A holder that is asked
again with an id it took follows H-A-364 (b): while the first act still runs the repeat joins it and gets the same
answer; after it ended the repeat answers the act's state as it stands, read afresh (the registry's publish answers
the transfer's parts held; a git verb answers the pull request as it is: merged, with its merge commit), which is
what the act answers when done; the act is never performed twice and no answer is kept in `taken`. A repeat for
which the state says the act never happened (the transfer is gone) is a refusal naming it. (2) The holder never
receives the caller's bound, so the taken key's TTL is a setting of each holder, `ask-taken-bound`, default 10
minutes (the largest caller's bound, the runner's ask-unanswered-bound), declared in the registry's and the forge
module's manifests; the node-engine keeps compose-ask-bound as H-A-364 (a) decided. §775's "the ask's bound" reads
so. P3, ADR 0668 §2, principle 11.
- **H-A-382 (a page by cursor: seen-again and each kept message, 00:16)** §779, T1. (1) The runner keeps nothing per
read and a bucket keeps no trace of a key's removed older message, so seen-again marks every message written after
the read's first page (the cursor carries that page's stream sequence), a key first created after the read began
included: no state per read in the runner's memory (the mesh owns the truth, P-NATS-native), and a reader replacing a
value it does not hold loses nothing. (2) A page hands each kept message in sequence order, a key's older kept
revisions included, each later revision of a key already handed marked seen-again by (1); only the first page may
ask last-per-key, and it does not: one way for every page (P3), as §779's acceptance says ("delivers each kept
message once"). P3, P-NATS-native, H-A-227.
- **H-A-383 (the node-engine's taken state names its cap; the baseline gains the word, 00:20)** E2. §820 gives every
holder's `taken` 8 MiB, and every kept store names its cap (§812, ADR 0595); the node-engine is no exception. Its
manifest declares taken {history 1, key-ttl, max-mib 8}. The word `state.max-mib` is one the engine reads already
(it writes every module's bucket cap from it), so mesh-sdk's baseline manifest reads (ADR 0614 §4) were incomplete:
the baseline gains `word:state.max-mib`, in the same group, and packaging/operator_keys_test.go follows the baseline.
The running engine on each node holds the new engine's build for a yes once, at this delivery, because its baseline
lacks the word: that hold is said to the operator before the delivery starts and answered then; it is the delivery
system doing what ADR 0614 asks, not a reason to leave the store uncapped (design-it-as-it-should-be).
P-failure-loud, §812, ADR 0614 §4.
- **H-A-384 (three node-engine lines the developer thought wrong, 00:24)** (1) §813 (4), H-A-357: the nft comment is a
label, the composition is the truth. Past 128 bytes the module list is cut with a count, `derived a,b,+4 tcp/443
scope any`, the whole list standing in the composition's filtering resource; the holder's rules reading (mesh-catalog
P1) reads that form; the compose never refuses for a label's length (a refusal would take the whole table down for a
shared port: not proportionate; P-failure-loud is served by the composition, not the label). (2) §813 (1), H-A-355:
the applied run is kept in the engine's have, the record of what it applied that it already writes to the bus (the
composition carries no run, as H-A-355 says); a restarted engine reads its have and never applies an older run again.
Memory is never the truth (P-mesh-owns-the-truth). (3) §813 (12), H-A-358: drain-bound raises the declaration's
form (the process's form rises as DMC-16 raised the build-step part's), so the delivery plan derives the ordering
edge node-engine (the reader of the new form) before the composer that writes it; the older engine's strict decode
stays strict. The edge is named in the milestone plan with this answer (design-it-as-it-should-be). P3,
P-mesh-owns-the-truth.
- **H-A-379 (addendum, 00:24)** The install ask carries no artifact or fingerprint, so the delivery SDK's check takes an
install's end without them, naming module, commit and the archive (mesh-delivery 1767d7b4).
- **H-A-385 (the gate's ledger entries: what each keeps, 00:34)** §809, GT1. (1) The head entry `<branch>.<source>`
keeps every field about the record as a whole: group, trace, acts, away, recovery, call, the pull requests' heads,
with its outcome, the movers' count and §810's times; nothing per module. (2) A mover entry
`<branch>.<source>.<module>` keeps that module's step outcomes per node (taken-in, built, checked, composed, wanted,
running-healthy, held, failed, away, each with why and event id) beside its pipeline, version, machines and running
times: §809 forbids an entry naming two modules, not a module's own detail, and delivery_status and show read them
there; the 64 KiB cap holds per entry. (3) The catalog half, when it moves, is a mover with an entry of its own
(`<branch>.<source>.mesh-catalog`, no node) and counts in the head's movers. (4) The month list (type held) stays
the way the ledger is listed; nothing changes it. One way per thing, P3; ADR 0595 §3.
- **H-A-386 (the keep's prefix is `_keep/`; the runtime's put answer carries the fill, 00:14)** Corrects H-A-380 (2):
the bus refuses a key-value key beginning with '.', so the keep lives under the reserved prefix `_keep/<key>`, the
underscore form the tool runner already uses for `_meta` (H-A-351); everything else of H-A-380 stands. DMC-14, which
H-A-350 answered and no build map carries: the tool runner's put answer (mesh/state.put) carries the entry's fill as
`{"revision": n, "fill": {"count", "bytes", "max_bytes"}}`, buscalls.Fill's form, which mesh-sdk's bundle-side prune
helper reads; built by the mesh-tools developer in H2 beside T1, tested in the runtime's state tests. P3, H-A-227.
- **H-A-387 (the artifact registry takes a publish whose reservation names no digest yet, 00:20)** Since H-A-373 the Go
build module publishes an archive with its reserved version during the build step, when the builder's reservation
still holds an empty digest for every artifact (the builder records digests only at the build's end); the
distribution module's holder (the seat oci-artifact-registry) accepts only an equal digest and says "for another
digest" of an empty one, so every archive-only publish since 00:05 was refused (twelve modules on three nodes, nftables
among them). A-490j decided this on 19 Nov for every holder: a publish is taken when the reserved artifact names that
digest or names none yet; commit ac2cd374 kept the strict check in the distribution holder against it. Fix in the
distribution module (mesh-catalog): holder.Reserved.Recorded takes an empty reserved digest as A-490j says, and its
refusal names the case it met ("reserved for digest X, not Y" or "no reservation of version V"), never "another
digest" for none. Test: a publish against a reservation with no digest yet is taken and marked; one against another
digest is refused naming both. The pack is reproducible (the same archive bytes at 0.1.0, 0.1.1 and the refused
publish). Recovery: rerun the build steps once the distribution module runs on novox. A-490j, P-failure-loud,
fix-a-bug-met-twice (the same symptom met 19 Nov).
- **H-A-388 (the have carries the applied run; the declaration's version is its form, 00:26)** E4b, E7b. (1) The
node-engine's have value gains `run` (the composition run's sequence the module was applied from) in
seats/node-engine/seat.json, generated, its form raised by one as H-A-352 (2) has it; the node-engine's developer
makes that K row in the mesh-contracts worktree (one commit citing this answer) since no other developer holds it
tonight, and tells the lead both heads. (2) The declaration carries no form beside its Version, so Version is its
form (one field per thing, P3): mesh-sdk's declaration gains the process's drain-bound under Version raised by one,
the composer (the assigner, mesh-delivery) writes that Version, the node-engine reads it and refuses a higher one as
it does today. The edge the delivery plan derives: the node-engine's manifest states the declaration Version it reads
(`declaration-reads`, beside its manifest reads of ADR 0614) and the composer's manifest the Version it writes
(`declaration-writes`); the plan orders the writer of a Version after every node's engine reads it. Built: the field
and the two manifest words in mesh-sdk (its developer), the engine's word and reader (mesh-host), the composer's word
and writer and the plan's edge (mesh-delivery, with A1-A6). Until the plan reads the edge, the milestone plan names
the order by hand: node-engine before the assigner. design-it-as-it-should-be, P3, ADR 0614.
- **H-A-389 (a validate step judging a candidate that changes a seat's form, 00:26)** The mesh-catalog developer found
the trunk validator refuses H2's delivery seat definitions (form, max-age: fields its format does not know), so a
pipeline's validate step, asked of the running assigner, cannot judge a candidate that raises a seat's form. Today
validate is an optional step (main.175 shows it so), so the H2 group's pipeline is not failed by it; the gap stands:
a candidate that changes mesh-delivery-sdk or a seat's form is judged by a validator built from the candidate's own
heads (ADR 0573 §2's rule for the check, carried to the validate step: the runner builds the validator tool from the
candidate like any artifact and asks it). Queued for H2b in mesh-delivery; the H2 group's validate verdicts are
read by hand once after its merge. design-it-as-it-should-be, ADR 0573 §2.
- **H-A-390 (§799's shared parts are the delivery SDK's developer's, 00:26)** BM1: the shared parts in
mesh-delivery-sdk's steprun (again-delays parsed into Puller.Delay, the third run's "fetch:"/"publish:" why,
HolderSilentOr's assigner call "with no seat call to learn why") are built by the mesh-delivery developer with its B
lines; the mesh-catalog developer builds the three build modules' settings and their use of the conformance kit
against the relayed head. One owner per file (P3).
- **H-A-391 (the gate's mover entries are found by an index on the record's key, 00:30)** §809, GT1. The ledger type
`mover` declares an index on its record's key in the gate's manifest `requires` (records' `indexes`), and
gate_record's join is one `mesh/ledger.query` on that index with the key `<branch>.<source>.>`; the head entry stays
per record (its count of movers and nothing per module), and nothing reads the revision's graph again. A mover's
instance path is `mover/<token of branch.source.module>` beside `record/<token of branch.source>`. P-NATS-native, P3.
- **H-A-392 (what H-A-374 changes in the node-engine, 00:30)** (a) Issue 339's rule stands: a directory the mesh never
made, found with another owner or mode, is used as found and said so; H-A-374 holds for a directory the engine's
record says the mesh made. (b) Two things, both read-only on the judgement: the verb composition with diff reads a
record-held directory's owner and mode on disk and names one that differs as drift; and the engine converges the
owner and mode of every record-held directory entry at its start and whenever it applies that module, one journal
line per directory set right, so a drift lasts at most until the engine's next start. An unchanged composition is
still not applied (idempotence keeps). Test: a record-held directory with another mode is set right at the engine's
start; one the mesh never made is left and said. P-mesh-owns-the-truth, issue 339.