Merge worktree/schedule-conversational-after into worktree/schedule-explicit-at

# Conflicts:
#	apps/web/tests/schedule-after.e2e.ts
#	packages/client/runtime/src/client/sessions/session.ts

packages/client/runtime/README.md

@@ -36,11 +36,15 @@ SlotsService gives the renderer separate bare observables for `useSessions` and
 
 `ConversationSnapshot.queue` is the Host's authoritative transient snapshot of `agent.inbox.nextTurn`; pending next-step steering stays outside this projection. Each row carries its `MessageId`, complete editable text when every content block is text, and a flattened preview. The Host derives whole `session/queue` snapshots from durable `agent/inbox/spliced` mutations and sends a baseline on reconnect; the message-local `agent/inbox/inserted`, `claimed`, and `discarded` notifications are not used to reconstruct this projection. `Session.updateQueue()` sends edit/remove operations through Host-side `Inbox.splice()` without optimistic client mutation, so the next Host snapshot is the sole visible commit and a claim race can surface `queue-item-not-found`.
 
-## The human transcript
+## Conversation assembly
 
-`ConversationSnapshot.nodes` is the human transcript, not the model surface. `TranscriptAdapter` projects the raw window in log order — every append-origin surface event (`isAppendSurfaceEvent`) at its own log position, plus one `CompactionSummaryNode` marker per landed compaction checkpoint — and never consults surface order. `SteeringHistory` replays the durable `agent/inbox/spliced` records in that window: a user-origin message claimed from `next-step` becomes a `SteeringMessageNode` when its matching `user/message` lands, a `next-turn` claim stays a user node, and non-user next-step input stays context. `ConversationSnapshot.turnEnds` maps each completed turn in that window to its `turn/end` seq, retaining turn completion independently from the transcript so presentation can require a real boundary before enabling an action. A landed compaction therefore keeps the conversation it shadowed on the model side: the marker reports where the model stopped seeing that history instead of erasing it. Model-only replacement copies stay out: a pruned `tool/result` and a regenerated `assistant/message` rewrite one node for the model and mark no boundary. A checkpoint is a `user/message` carrying the compaction seam's plugin source that **replaced** a surface range; an appending plugin-sourced `user/message` is injected context, not a compaction. Each context node also carries a `provenance` view with the producer role and name: `contextProvenance()` reads the durable source alone to decide whether the row is an `inject` or a cross-session `recall`, and to name its producer from the instruction paths, referenced session titles, or plugin id that source already records. The client holds no table of plugin ids, so a renamed or newly mounted producer stays identifiable without a client release and a resumed or foreign log projects exactly like a live one; a source with no readable kind degrades to an unnamed injection. Beside it, `contextForm()` reads the producer-declared `ContextForm` — the second, independent axis: `kind` says who produced the context, `form` says what shape of information it is, so several producers may share one form. A form this UI version does not present projects as null and renders opaque. The adapter's plugin literal is pinned to the Service Definition's declaration by a type-only import of the cordis-free [`dsh-compact/checkpoint`](../../compact/compact/README.md) leaf, so renaming it there fails `tsc` here; a **value** import of the package would fail the client purity gate, and the package **root** is unreachable even as a type (it reaches `dsh-session`'s root, whose `Context` merge collides the host `sessions` with this program's).
+Each `Session` gives its contiguous event window to a `ConversationNodeAssembler`. Plugins register business Definitions that map one event to a stable `{kind, id}`, create State at the unique start event, fold correlated updates, and build final nodes for registered view targets. The assembler owns the Context index, read-only predecessor lookup, and a reference-stable Turn/Step Location index. A live append evaluates each Definition once and updates only the matched Context; loading an older page preserves existing Context and node identities, matches only the newly prepended events, and replays Contexts whose predecessor or Location facts changed. Full replacement is reserved for open, resync, and gap repair.
 
-Because the projection is log-ordered, the node array is seq-monotonic by construction: log-only `command/run` / `command/done` nodes splice in by seq, `Session` merges interrupted frozen nodes by their fractional seqs, and a window whose checkpoint cites a shadowed range outside it renders the marker with nothing logged. The marker's summary text, replaced-item count, and estimated shadowed-token count come from the checkpoint's cited `compact/summary` event; a window cut that left that event outside makes those fields unavailable, and a later page that supplies it resolves them. `CommandNode.outcome.sourceEventSeq` preserves a successful command's explicit reference to that summary event, allowing the presentation layer to pair `/compact` with its checkpoint without parsing settlement copy or assuming the two rows are adjacent. Performance contract: one append materializes at most one node and copies the projection only when it adds that node; an event that changes no node keeps the previous array reference (a chunk storm costs nothing), and unchanged nodes keep their object identity.
+Definition authors keep matching local to the current event, give every correlated event a stable business id, and make updates replayable by log `seq`; renderers consume final Node data and constrained Location values rather than scanning Session or Chat collections. The [Conversation Node cookbook](../../../docs/cookbook/adding-a-conversation-node.md) gives the complete registration and pagination path.
+
+`ui-conversation` registers the built-in Chat Definitions and the keyed Chat snapshot builder. Append-origin user, assistant, and Tool results remain the human record; model-only replacement copies stay out, except that a compaction checkpoint becomes its own marker and resolves missing summary provenance when an older page supplies it. Durable inbox splice Contexts classify next-step user messages as steering without making inbox state a Session special case. Context messages retain producer provenance and form. StatsLine reads `ConversationSnapshot.chat.legacy.nodes`, while Session mirrors that legacy slice into the top-level `nodes`, `partial`, and `runningCalls` public compatibility fields without running a second business fold. Trajectory consumes neither compatibility surface; its activated `session-history` inspection keeps an independent fold until it gains its own registered target.
+
+The Chat builder keeps one mutable keyed store per Session. Content updates notify only the affected node key, structural changes rebuild order and Location membership, and a prepend adds rows without replacing existing keyed values. Assistant chunks update Definition State for every event but request at most one materialization per animation frame; final messages and Turn/Step closure publish immediately. See the [client Tool presentation decision](../../../.agents/notes/implemented/architecture/2026-08-08-client-tool-presentation-ownership.md).
 
 ## Request inspection
 
@@ -48,7 +52,7 @@ Because the projection is log-ordered, the node array is seq-monotonic by constr
 
 ## Code Mode child-call tree
 
-Every `ToolCallBlock` recursively owns its children through `subCalls`, in start order. Runtime's `ToolCallTree` privately maintains the parent-callId-to-children index: a `tool/code-dispatch-start` event lands as a `RunningToolCall`, and the matching `tool/code-dispatch` settlement replaces it in place with a `ToolResultNode` whose `callTime` comes from the paired start. When the start fell outside the replay window, the settlement appends directly with `callTime: null`; Runtime never fabricates a zero duration. Live mux frames and history replay share this fold and tree projection, and child calls never become independent roots in transcript `nodes`. A child update copies only its ancestor path to the owning root; unchanged siblings and other roots retain object identity. Wire or history edges that would introduce a cycle or exceed the fixed 256-call recursive-depth safety limit are consumed without mutating the tree, so the rest of the session remains renderable.
+Every `ToolCallBlock` recursively owns its children through `subCalls`, in start order. Chat's Tool Definition correlates root calls and results by call id, folds Code Dispatch start/settlement records into that root Context, and projects one keyed recursive tree; child calls never become independent Chat roots. When a start falls outside the loaded window, its settlement remains renderable with `callTime: null`. A child update copies only its ancestor path, so unchanged siblings retain object identity. Edges that introduce a cycle or exceed the fixed 256-call depth limit are consumed without mutating the tree. The separate Trajectory history fold still uses Runtime's `ToolCallTree` over the same nested data contract.
 
 ## Session title projection
 
@@ -56,7 +60,7 @@ Every `ToolCallBlock` recursively owns its children through `subCalls`, in start
 
 ## Model retry projection
 
-The Session object validates plugin-owned, provider-routed `llm/retry` payloads at the event wire boundary against the producer's complete field contract, including timer, integer, status, provider-delay, and non-empty diagnostic bounds. A valid event removes the matching failed step's streaming partial and inserts a durable retry notice at the event's sequence position. The notice is `scheduled` until a following retry turn starts; an aborted or disposed source turn marks it `cancelled`, while the retry turn marks it `started`. Normal-mode notices carry their finite maximum; always-mode notices remain explicitly unbounded. A terminal `turn/end` error without a retry projects one `turn-error` node from its durable message and optional code; AUTH projections replace provider copy that may echo credential fragments with `API key is invalid`, while the raw diagnostic remains in the session log. A retried failure keeps only the retry notice for that attempt. Window rebuild and history replay apply the same projection, so refresh neither resurrects discarded chunks nor loses terminal failure feedback. Visible unfinalized output is frozen as an interrupted assistant node beside the terminal error.
+The Host-owned LLM retry invariant validates provider-routed `llm/retry` and `llm/retry-started` records at the durable append boundary, including their identity, ordering, timer, integer, status, provider-delay, and non-empty diagnostic contracts. In the client, the Retry, Assistant, and Turn Error Definitions fold those records with Assistant and Turn/Step events: a failed step's streaming partial is removed and a durable retry notice appears at the retry event's sequence position. The notice is `scheduled` until the matching started record arrives; closing its owning Step or Turn first marks it `cancelled`, while the started record marks it `started`. Normal-mode notices carry their finite maximum; always-mode notices remain explicitly unbounded. A terminal `turn/end` error without a retry projects one `turn-error` node from its durable message and optional code; AUTH projections replace provider copy that may echo credential fragments with `API key is invalid`, while the raw diagnostic remains in the session log. A retried failure keeps only the retry notice for that attempt. Window rebuild and history replay use the same Definitions, so refresh neither resurrects discarded chunks nor loses terminal failure feedback. Visible unfinalized output is frozen as an interrupted Assistant node beside the terminal error.
 
 ## Session forking