Keep conversation turns distinct and history stable (#7990)

* fix(app): keep autonomous turns distinct and readable

Hidden system prompts merged autonomous work into the previous response and reopened completed turns. Preserve canonical turn boundaries and expose their triggers, while making received messages attributable and expanded turn details readable in one step.

* fix(timeline): mount history at actual run boundaries

* fix(timeline): load and retain complete available turns

* fix(app): summarize actions and preserve conversation events

* test(timeline): verify whole-turn paging stays responsive

* fix(app): group reasoning according to thinking preference

* test(app): align history journeys with whole-turn loading

docs/timeline-sync.md

@@ -13,7 +13,7 @@ and rebuilds the projection when an agent resumes.
 The invariants are:
 
 > A continuously subscribed client applies every committed row in order. Opening or resuming an
-> agent establishes the daemon's current tail in one bounded request, with older history reachable
+> agent establishes the daemon's current tail in one complete-turn request, with older history reachable
 > through backward pagination.
 
 Tool output is bounded before it enters either delivery path. Canonical shell tool output is sliced
@@ -44,7 +44,20 @@ recovery is complete only when `hasNewer: false`.
 
 Initialization timeouts guard lack of catch-up progress, not the full multi-page sync. A successful page that queues the next `after` page refreshes the watchdog.
 
-Opening or resuming an agent fetches one bounded latest tail page. Older history remains
+Opening or resuming an agent fetches the latest available turns with a soft target of 40
+projected items. Backward pages and prompt-jump windows also include whole turns, including
+all currently available rows of an active run. There is no per-turn cap: a large run can
+exceed the target. Local history mounting uses the same boundaries so revealing older rows
+does not change the collapse state of a turn already displayed.
+
+`wholeTurns` is optional for existing SDK callers. The app requires
+`features.wholeTurnTimeline` once at the client fetch boundary; unsupported hosts receive an
+update-host error. Forward catch-up stays incremental, but its reset snapshot honors the
+whole-turn request. Source coverage still includes interleaved tool updates.
+
+Canonical IDs define runs. With missing IDs, human prompts, heartbeats and typed received
+agent messages each start a fallback turn. This deterministic fallback is shared by paging
+and presentation; it cannot recover historical batching or steering. Older history remains
 user-driven by scrolling upward.
 
 A failed catch-up or subscription reconcile retries on its own, doubling from 1s to a 30s ceiling.
@@ -149,12 +162,13 @@ range describes those exact items: `startSeq` drives older pagination and `endSe
 catch-up. The owner requests `after endSeq`, and requests `before startSeq` when the user loads older
 history. Code outside the owner does not distinguish cached and network timelines.
 
-The first resume request is bounded. If it reports more newer history, fetch one latest bounded tail
+The first resume request is bounded. If it reports more newer history, fetch one latest complete-turn tail
 instead of replaying every missed page. Live gap recovery still pages forward until current.
 
-If the canonical window exceeds the cache item limit, contains a discontiguous retained range, has a
+If the canonical window exceeds the cache item target, contains a discontiguous retained range, has a
 live head, or includes presentation data the cache cannot encode losslessly, persistence drops the
-range and keeps a display-only tail. Never slice items while retaining the pre-slice range; that
+range and keeps a display-only tail of complete available turns. The item target is soft;
+the byte budget still evicts entire cache rows. Never slice items while retaining the pre-slice range; that
 falsely certifies discarded source rows. A display-only row paints without granting synchronization
 authority, so the owner uses the ordinary bounded `tail` bootstrap.
 
@@ -238,11 +252,11 @@ foreground control ownership remains a separate daemon concern. Cancellation req
 with that record rather than in a React component, so an old request cannot clear a newer one. Submissions
 remain a separate pre-turn registry and retire on canonical acknowledgement.
 
-Canonical turns and visible responses are different boundaries. Agent-originated prompts and lifecycle
-notifications project into tool rows, so they do not create a human-message boundary. Legacy system
-prompts, including schedule injections, remain hidden. One visible response can span several canonical
-turns; layout and copy group that response together while lifecycle, timing, tool sequences, and exact
-fork positions retain the canonical `turnId` boundaries.
+Every canonical run owns a visible turn, including heartbeats and agent notifications.
+Notifications delivered during a run stay inside it; batched notifications share the run
+that starts or accepts them. Received messages retain sender attribution, while lifecycle
+notifications remain visible events. Collapse, timing and locally mounted history share
+these boundaries.
 
 Agent-message XML belongs to the daemon's `agent/agent-messages` boundary. Prompt delivery encodes
 provenance; acceptance, provider echoes, import, and replay share its projection. The delivery ID is