03-DESIGN/01-to-be/25-the-bus-on-nats.md

@@ -16,6 +16,7 @@ decisions:
   - "ADR 0619"
   - "ADR 0615"
   - "ADR 0621"
+  - "ADR 0622"
   - "ADR 0614"
   - "ADR 0613"
   - "ADR 0612"
@@ -752,13 +753,18 @@ vault as an own secret: JetStream create and update under modules' names only (s
   ADR 0621). The four
   foundation parts that read each other (the grants part, the tool runner, the indexer and the assigner) stop
   reading the indexer's and the assigner's bus state in the milestone of ADR 0612, with the other 19 readers of that
-  state (ADR 0615 §1). **Each owner serves them as tools of its own module**, the indexer `manifests` and
-  `modules`, the assigner `placement`, `seat-holders` and `wants`, and declares `runs-as`, so its tools bundle runs in
+  state (ADR 0615 §1). **Each owner serves them as tools of its own module**, the indexer `manifests`,
+  `modules` and `verdicts` (ADR 0622 §1), the assigner `placement`, `seat-holders` and `wants`, and declares `runs-as`, so its tools bundle runs in
   a runtime of its own on the module's own bus account, never in the node's tool runner, and the tool runner never
   asks through itself (ADR 0621 §1). No new manifest form is added: a process still answers no tool call or seat verb
   (ADR 0436 §3). **Each verb is granted to exactly its readers** (ADR 0621 §2, as ADR 0615 §3 lists them): `placement`
   to the grants part, the tool runner and the indexer; `seat-holders` to the tool runner; `wants` to the grants part;
-  `manifests` and `modules` to the grants part, the tool runner and the assigner. A reader declares each in `invokes`
+  `manifests` and `modules` to the grants part, the tool runner and the assigner; `verdicts` to the grants part
+  alone, its `callers` list holding the grants part only (ADR 0622 §1). These six are the first-view tools.
+  `verdicts` follows the first-view rules whole (ADR 0622 §2, under ADR 0615 §4 to §6): its answer carries the
+  revision, the grants part then follows `manifest-accepted` and `manifest-refused` and asks once more on a gap in
+  their revisions, it raises the condition naming the indexer and `verdicts` past the bound and never reads the
+  indexer's state, and `grants_first_views` and `grants_first_view_again` cover it. A reader declares each in `invokes`
   as `<module>.<tool>` only where it is granted, which is the grant the grants part composes as for any tool, and
   each tool checks the proved caller against a `callers` list holding exactly its readers and refuses any other, the
   owner itself included. **The judge of the milestone is its own group's** (ADR 0621 §4): since no field is added,
@@ -815,6 +821,21 @@ vault as an own secret: JetStream create and update under modules' names only (s
   part and the tool runner their first views while the bus module still waits on it. The order of owner before reader is kept by the module graph's served-verb edge for the 19 readers (ADR 0614 §2), and,
   for the four, whose verb edges lie on cycles the graph leaves out (ADR 0614 §3), by the owner's answer at once and
   the reader's wait (ADR 0615 §2).
+- **The other 19 readers ask the owners' read tools, one per bucket they read** (ADR 0615 §1, §2). Beside the six
+  first-view tools above, each owner serves a read tool for every bucket of its state one of the 19 reads today,
+  named `indexer_read_<bucket>` or `assigner_read_<bucket>` and marked as a read: fifteen. The indexer's nine:
+  `indexer_read_manifests`, `indexer_read_modules`, `indexer_read_predefined`, `indexer_read_config_candidates`,
+  `indexer_read_candidates`, `indexer_read_packages`, `indexer_read_graph`, `indexer_read_heard` and
+  `indexer_read_taken`. The assigner's six: `assigner_read_placement`, `assigner_read_seat_holders`,
+  `assigner_read_wants`, `assigner_read_findings`, `assigner_read_asks` and `assigner_read_differences`. These are
+  the nine buckets of the indexer and the six of the assigner the 19 read (§ 12.6), `modules`, `wants` and
+  `placement` among them for the status view; the indexer's tenth, `verdicts`, is read by the grants part alone,
+  through its first-view tool. **They are ordinary module tools**: granted, as any tool is, to every
+  reader that declares it in `invokes`, and callable by a session; the `callers` list of ADR 0621 §2 and ADR 0622 §1 covers only the six
+  first-view tools. **The status view keeps all its tools**, their names, arguments and answers
+  unchanged, its `indexer_*` and `assigner_*` tools included: as a reader it answers them by asking the owners' read
+  tools instead of reading their buckets. Checked live after the delivery: a session's call of each status view tool
+  answers, and `mesh_search` finds each owner's read tool by its bucket's words.
 - **Only the owner reaches it** (ADR 0596 §2). The membership grants the owner's account its own entries and no
   other module's: the new form has no readers. What another module needs of that data it asks through the owner's
   tools or seat verbs, and what the owner lets others hear is its declared events, subscribed to through
@@ -969,12 +990,17 @@ this milestone go on their old sections too (§ 12.2; design 51 *The move*; desi
 
 **Readers move after their owners** (ADR 0596 §3). Today 195 `reads` entries in 37 modules read another module's
 bus state (counted 2026-11-03 02:00). **A correction of fact**: the indexer's and the assigner's state, counted again
-from every manifest on main later that day, is read by 23 modules, 9 buckets of the indexer and 6 of the assigner,
+from every manifest on main later that day, is read by 23 modules, 10 buckets of the indexer (`verdicts`, read by the grants part, among them) and 6 of the assigner,
 not by the 19 the first plan counted; and 151 state entries were counted, not 158 (the template table of § 12.1 keeps
-the first count, which chose the templates). The owners first gain their events and first-view tools: the indexer
-one per bucket its readers read, the assigner the same. Then all 23 readers move in the milestone of ADR 0612, the
-four foundation readers included, by the first view of § 12.2 (ADR 0615 §1); their order is the module graph's, not
-a review reading an owner delivered (ADR 0615 §2). The judge still refuses a moved module's subscription to an event
+the first count, which chose the templates). The owners first gain their events, first-view tools and read tools: the
+indexer a read tool per bucket its readers read, the assigner the same (§ 12.2 lists them). Then all 23 readers move in the milestone of ADR 0612, the
+four foundation readers included, by the first view of § 12.2 (ADR 0615 §1), the grants part's of the indexer's
+`verdicts` by its first-view tool `verdicts` (ADR 0622) and then `manifest-accepted` and `manifest-refused`; their order is the module graph's, not
+a review reading an owner delivered (ADR 0615 §2). The status view's move keeps every tool it serves, each answered by
+asking the owners' read tools of § 12.2. **The node-engine's reads granted in code stay**: the grants part's code
+grants each node's node-engine, with no manifest declaring them (`CandidateReads` and the composed reads beside it),
+the indexer's `candidates` and `verdicts` and its own node's keys of the assigner's `wants`. They are not among the
+23 readers of ADR 0615 §1 and move in the next milestone, queued on issue 1966. The judge still refuses a moved module's subscription to an event
 its owner does not declare. Every other reader moves when its module is next touched. These moves are tracked by
 issue 1966.
 
@@ -986,10 +1012,10 @@ issue 1966.
 | The bus module creates, and the binding follows | live, by tool call: a module asking for a state, a log, objects and a ledger in `requires` receives each binding in its membership, written after `created`, never before; `show` gives each entry's template, settings and fill; a module calling `create` is refused |
 | The bus module runs whole | one integration test in the nats module's repository that starts the bus module whole against the real bus server: a verdict, then `created`, then the binding in the membership; a restart that catches up a verdict given while it was down; a bucket whose settings differ raises `drifted`, the seat's `apply` of a larger cap clears it with `drift-cleared`; a refused create raises its condition and hands no binding; `recreate` and `drop` without the operator's warrant are refused, and with it act after a stream snapshot; an orphan is listed, raises `orphaned`, and goes only by `drop`; each template at a small cap refuses, events drop the oldest, and `full` is raised at the raise mark and cleared only below the clear mark; `recreated` rewrites the binding and `dropped` removes it; at start the first view from `indexer_bus_state` creates what a verdict given while it was down asked; the older remove tools refuse a new-form entry, naming `drop`'s tool |
 | A module asking bus state waits for the bus module | the module graph's table test: a manifest with a ledger entry has a requirement edge to the module providing `bus-state`, one without has none (ADR 0614 §5) |
-| A foundation part's first view | unit tests of each owner and reader (ADR 0615's How it is checked), as table cases, and these cases of the whole-module integration tests that exist today (mesh-delivery is over its integration budget, so no new test function): the indexer's (`cmd/mesh-delivery-indexer`, its one whole-module test, which the milestone's packages test merges into), the assigner's (`cmd/mesh-delivery-assigner`) and the grants part's (`cmd/mesh-grants`), and the tool runner's in mesh-tools: the owner answers its verb at once from what it holds, with its revision, before its own first view; the reader asks once at start, then only on a gap in revisions; a reader whose owner is down raises its condition only past the bound and never reads the owner's state, and the owner's answer clears it; an empty answer at revision 0 drops no login; an owner's event carries the changed value whole, and a reader applies a put and a delete from the events alone and never asks the verb; a revision lower than the one held makes the reader ask its first view again; an existing waiter on `manifest-accepted` or on the assigner's `put` does not wake on the new families; the read tool shows the revision each reader holds, and the re-read tool asks once and records the hand act; in the
+| A foundation part's first view | unit tests of each owner and reader (ADR 0615's How it is checked), as table cases, and these cases of the whole-module integration tests that exist today (mesh-delivery is over its integration budget, so no new test function): the indexer's (`modules/mesh-delivery-indexer`, its one whole-module test, which the milestone's packages test merges into), the assigner's (`modules/mesh-delivery-assigner`) and the grants part's (`modules/mesh-grants`), and the tool runner's in mesh-tools: the owner answers its verb at once from what it holds, with its revision, before its own first view; the reader asks once at start, then only on a gap in revisions; a reader whose owner is down raises its condition only past the bound and never reads the owner's state, and the owner's answer clears it; an empty answer at revision 0 drops no login; an owner's event carries the changed value whole, and a reader applies a put and a delete from the events alone and never asks the verb; a revision lower than the one held makes the reader ask its first view again; an existing waiter on `manifest-accepted` or on the assigner's `put` does not wake on the new families; for the indexer's `verdicts` (ADR 0622), called by the grants part while the node's tool runner is stopped, it answers, and from any other account, the indexer's own included, it is refused; a missed revision on `manifest-refused` makes the grants part ask `verdicts` once more; the grants part and the indexer started at the same moment, `verdicts` answers within the bound with its revision; with the indexer down past the bound the grants part raises the condition naming the indexer and `verdicts` and never reads the indexer's state, and the answer clears it; the read tool shows the revision each reader holds, and the re-read tool asks once and records the hand act; in the
 | The owners' first views are their own tools, in a runtime of their own (ADR 0621) | cases of the indexer's and the assigner's whole-module integration tests: each first-view tool, called from a reader's account while the node's tool runner is stopped, answers (§1); each answers its granted readers and refuses every other account, the owner's own included (§2); the two runtimes started at the same moment, each asking the other's first view at start, both answer within the first-view bound with their revisions (§3); and the milestone's group passes its merge check with no change to the judge's known fields for these verbs (§4) |
 assigner's whole-module test, the assigner starts before any new-form bus state exists, with no binding of the bus
-module, and the indexer, the grants part and the tool runner each get their first view of it. Live, after the delivery: the indexer's count of `reads` entries on the indexer's and the assigner's state reads 0 |
+module, and the indexer, the grants part and the tool runner each get their first view of it. Live, after the delivery: the indexer's `indexer_old_sections`, which the milestone extends to count the `reads` entries on each owner's state, reads 0 declared `reads` entries on the indexer's state and 0 on the assigner's, and says the node-engine's reads granted in the grants part's code remain (§ 12.6); and each owner's read tool of § 12.2 answers what its bucket holds, each status view tool answers a session's call as before the move, and `mesh_search` finds each owner's read tool |
 | A foundation part's first view comes within its bound | live, after the delivery: each reader's read tool (`grants_first_views`, the mesh MCP server's `mesh_first_views`, `indexer_first_views`, `assigner_first_views`) shows, for the indexer's and the assigner's verbs, the time from its ask to the owner's answer under the two-minute bound, and no first-view condition raised by the delivery |
 | The bus module's first view | cases in the nats module's one integration test: a replayed verdict older than the first view is not applied; an `indexer_bus_state` that does not answer turns the health unhealthy naming it and is asked again on the backoff; a set-aside manifest missing from the first view keeps its `asked` key, an empty `manifest-accepted` removes it; a deleted module's entry, after its empty `manifest-accepted`, is listed by `orphans` and raises `orphaned`, and a set-aside module's entry is neither; `nats_first_view_again` asks at once and records the hand act |
 | `manifest-accepted` is never lost silently | unit tests of the indexer: a failed publish is published again, and one still failing writes the finding; `indexer_say_accepted` says it again for one module from the verdict held, and refuses a module with no accepted verdict |

03-DESIGN/01-to-be/51-the-registries-keep-what-is-named.md

@@ -9,7 +9,8 @@ code:
   - mesh-delivery modules/mesh-delivery-builder (versions, tag asks), the forge bridge, the indexer, the checker, the assigner
   - mesh-sdk (the bus helper for sending in parts)
   - mesh-catalog modules/docker (image pruning)
-  - mesh-host internal/apply (every new image of a module pulled before any of its containers is replaced)
+  - mesh-host internal/apply (every new image of a module pulled before any of its containers is replaced; a failed pull applied again on the status view's `settled` for the store's holder, and at the node-engine's start), internal/compose (each routed image reference written with its seat and that seat's holder)
+  - mesh-delivery modules/mesh-grants (what a node-engine hears: the status view's `settled`)
 updated: 2026-11-04
 decisions:
   - "ADR 0605"
@@ -194,8 +195,32 @@ the verb.
   composition, on the node-engine's verb `compose` naming the module (by hand), or when its liveness judge finds a
   declared resource missing or down; only a failure of the bus itself is tried again on a clock (a bounded backoff,
   at most four times before the have says failed). So another module's apply whose pull falls in that window stays
-  failed, its running containers kept, until one of those: its have says failed naming the image, and `compose`
-  naming the module is its recovery.
+  failed, its running containers kept, until one of those: its have says failed naming the image.
+  **So a pull that falls in the swap is tried again when the store's holder has settled, never on a timer.** Neither
+  the condition keeper's `condition-cleared` (ADR 0564 §3) nor the status view's `delivered` comes in the swap: a
+  finding of the holder held during its own delivery raises no condition and so no clear (ADR 0564 §13 (c)), and
+  `delivered` waits for every have to equal its want, which the failed have itself blocks (design 47). The status
+  view's `settled`, said once for a module on a node at one composition when its have reaches its want running and
+  healthy (design 47), is blocked by nothing else. `distribution` claims a seat held once for the mesh and runs on
+  one node, so the node-engine on every node hears `settled` for the store's holder, on whichever node the holder
+  runs, and then applies once more each module on its own node whose have says failed naming an image of that store,
+  said in its journal. **No `settled` is missed in a race**: the node-engine keeps in memory the last `settled` it heard
+  for each holder since its own start, never reading the status view's state (ADR 0596). When it records a pull
+  failed against a store, it compares the pull's start with that event's time: only when a `settled` for that
+  holder was heard after this pull started does it apply again at once. That is the one retry, once per `settled`: a
+  second failure after the same `settled`, or the at-once retry failing again, starts no further apply. It finds the holder by the seat the failed image came through, read from its
+  declaration, never by a module's name: **the composition writes, beside each image reference it routes, the seat the
+  reference came through and that seat's holder**, taken from the holdings the composition already has, so the retry
+  matches `settled` by the holder its declaration names. A change of the seat's holder composes the declaration
+  again, so the name it carries is the holder of now. The seat and the holder are not part of what a container is
+  (they are not in its digest), so a declaration composed again for them replaces no container. **The seam that changes is the grants part's list of what a node-engine
+  hears** (`NodeAgentHears`; the node-engine declares no `consumes` of its own): it gains one entry, the status
+  view's `settled` family, which the node-engine filters for the holder of the seat its failed image names. **At its
+  own start** the node-engine also applies once more every have that failed naming an image of a store, so a restart
+  or an absence across the swap misses nothing. A module already being applied is not applied a second time; a retry
+  whose pull fails again says failed again and starts no further apply. A `settled` of any other module starts
+  nothing, and no clock starts an apply. `compose` naming the module stays, but only as the hand recovery of a pull
+  still failed after that, recorded as a hand act; it is no step of the delivery.
 - **The package seats' read sides need no credential** for the mesh's own packages: the forge serves the mesh owner's
   packages to anonymous reads, so the provisions `npm-package-registry` and `go-package-registry` carry no
   credential (form none), and mesh-build-node, which held a pair for both writing and installing, requires none (the
@@ -463,6 +488,21 @@ test).
 - **Node-engine, unit** (a case of its apply test, with the runner it already fakes): a module of two containers
   whose second new image cannot be pulled replaces neither container, says the image, and keeps both running; with
   both images pullable, both are pulled before the first container is set aside.
+- **Node-engine, unit** (cases of its apply test, with the runner, the bus and the clock it already fakes): a
+  module whose pull failed, its have failed naming an image of a store, is not applied again with the clock moved
+  past the bus backoff's four tries and no event heard; it is applied once more, and its new image pulled, on a
+  `settled` for the holder of the seat its image came through, a `settled` said for the holder on another node
+  included; a pull that fails after a `settled` for its holder heard since that pull started is applied again at
+  once, while a `settled` heard before the pull started applies nothing at once, a second failure after the same
+  `settled` applies nothing, and a race retry that fails again starts no further apply; a
+  `settled` of another module applies nothing; at the
+  node-engine's start it is applied once more with no event; a module being applied when `settled` arrives is not
+  applied twice; a retry whose pull fails again says failed again and starts no further apply.
+- **Node-engine's composition, unit** (cases of its composition test): a declaration's routed image reference names
+  the seat it came through and that seat's holder; a change of the seat's holder composes the declaration again with
+  the new holder named and recreates no container.
+- **The grants part, unit** (a case of its composition test): a node-engine's subscriptions hold the status view's
+  `settled` family beside the backup holder's word, and nothing more.
 - **Checker.** A changed build module invoking a verb no running holder serves is not run, and the verdict names the
   modules composed from their manifests instead, and the verb.
 - **Vault.** Reverting the milestone's merge mints the build modules' credentials again; the bootstrap's one-off
@@ -474,7 +514,9 @@ test).
   shown by `builder_versions`; a node runs a module from a reference in the new form; a node-engine away across the
   milestone wakes and applies the new node-engine; the node-engine's apply journal for `distribution` on each node
   shows every new image pulled before its first container was replaced, and no node's pull of any module failed
-  during the delivery (each node's haves, none failed naming an image); mesh-build-node installs from a lock file
+  during the delivery (each node's haves, none failed naming an image), or, where one fell in the swap, its have
+  turned healthy after the status view's `settled` for the store's holder, the node-engine's journal naming that
+  `settled`, and no `compose` was called for it; mesh-build-node installs from a lock file
   naming the registry's public root with no credential. Go's `direct`: its removal is named on the milestone's issue,
   and the change that removes it adds to the Go build module's recipe test a case that refuses `direct` in the fetch.