feat(chat-history): end the transcript with a live agent status trail

[Lando801]

Problem

A running round told the user almost nothing, and it told it twice.

PlanningFooter rendered a row saying "Planning next step..." / "Agent is
typing...", but usePlanningIndicator hides that row for a second after every
store mutation — so during a tool batch, which is exactly when a long round is
least legible, the transcript simply ended. Nothing on screen said how long the
round had been going, and the finished turn's "Agent worked for X" bar only
appeared once the round was already over.

Nothing carried the session's identity either: the transcript stopped at the
last event rather than at the agent that produced it.

Solution

One row now closes the transcript, for the whole session:

◼  Agent working for 1h 20m 26s · Agent is typing...
◼  Waiting for your reply
◼  Agent is idle
◼  2 running tasks · Agent is idle
◼  Last refreshed 5 minutes ago

• AgentStatusTrail replaces PlanningFooter. The activity phrase became
  this line's second segment rather than a second line, and it keeps its exact
  behaviour — same usePlanningIndicator count, same stable variant index, same
  agentTyping/compacting selection (pickPlanningVariant was extracted, not
  rewritten). The duration is the row's anchor so the row cannot flicker when
  the phrase hides itself between beats.
• "Agent working for X" is the present-tense twin of the finished turn's
  "Agent worked for X" bar, anchored to the same ChatGroupMeta.startMs, so the
  live line and the line it becomes are the same sentence over the same span.
• Phase mirrors the sidebar dot. resolveTrailPhase applies the same
  precedence buildSessionMenuItem uses — asking above running above resting —
  reading the session's own status via isSessionInProgress /
  isSessionPendingAsking, with liveness through the existing
  resolveTailTurnAgentWorking. The trail, the tail-turn collapse phase and the
  sidebar dot now share one rule.
• A quiet round stops reading as live. No transcript activity for five
  minutes and the trail drops to resting whatever the status says: nothing
  writes a terminal status for an imported transcript when that process exits,
  and a dropped agent:complete leaves the native atom on running. Either way
  the old behaviour would have been a counter climbing forever next to work that
  had stopped.
• Imported transcripts do not claim agent state. ORGII does not run an
  imported Claude Code / Codex / Cursor agent and cannot know it is idle, so
  those sessions report DataSourceConfig.lastScannedAt — the same timestamp
  the Runtime scanning panel shows — instead of "Agent is idle".
• The mark is the session's own harness, via SessionIdentityIcon (the
  sidebar row's projection), breathing on a slow agent-pulse while running and
  parked at that animation's own low point when resting, so switching phases
  changes whether it moves, never where it sits.

Live context tokens and the background-task count are deliberately not on the
running line: the composer's context pill and ActiveProcesses already carry
them. Running tasks do appear while resting, where "Agent is idle" beside two
live shells would be false.

Potential risks

• The row is now always present once a session has a turn, where the old
  planning row came and went. That is a persistent ~36px row at the end of every
  transcript, including historical and imported ones. Visual judgement call; the
  cheapest dial is dropping the resting label to icon-only.
• Not seen running. ORGII is Tauri-only with no browser preview, so nothing
  here is confirmed against real pixels — in particular whether text-2 plus a
  long planning phrase reads well at the 800px measure. Everything below is
  compiler- and test-level evidence.
• PlanningFooter is deleted. Its only consumer was this footer slot
  (verified by grep), but any out-of-tree import would break.
• Locale coverage is mechanical. The seven planning.statusTrail keys were
  authored for all 13 locales without a native-speaker review.
• Timer churn is confined to the leaf row: the state object carries
  startedAtMs, not a ticking value, so ChatHistoryList's memo is not
  invalidated once a second. A 1s clock runs only while running, a 30s clock only
  while showing "last refreshed".
• Scoped surfaces (subagent cells, cloud member-turn runners) still supply
  their own liveness and suppress the token/asking paths, since they have no
  sidebar row to agree with. Not exercised beyond unit coverage.

Verification

Commands run, all on this branch's content:

• npx tsc --noEmit --pretty false — 0 errors (in the working checkout).
  An isolated worktree typecheck at this commit timed out after 10 minutes,
  a known stall in this repo; instead I diffed the one file that had drifted
  against origin/develop and confirmed by grep that no PlanningFooter
  reference survives and that the rebased barrel still exports everything its
  two importers (ChatHistoryList, ChatHistoryListTypes) pull from it.
• npx eslint src/engines/ChatPanel/ — clean.
• npm run check:circular — no circular dependencies across 6407 modules.
• npx vitest run src/engines/ChatPanel/ src/i18n src/util/session src/store/session src/scaffold/NavigationSidebar — 272 files / 1919 tests passed.

Regressions reproduced before fixing, not assumed:

• Reverting the footer-slot gate to planningIndicatorCount > 0 alone fails 2
  tests in ChatHistoryListStatusTrail.test.ts; restoring it passes.
• Removing the stale term from resolveTrailPhase fails 3 tests across
  agentStatusTrailMath.test.ts and useAgentStatusTrail.test.ts; restoring
  passes.

New coverage: 44 tests over four files — duration formatting, activity and
phase precedence (including that waiting_for_user is itself an in-progress
status and must not read as running), the quiet-window timer arming and
re-arming under fake timers, the composed line under a real renderer, and the
footer slot surviving in both the virtualized and static render paths.

Also verified: .animate-agent-pulse, its @keyframes, and
.motion-reduce\:animate-none are actually emitted by compiling the Tailwind
config; every locale's planning.statusTrail key set matches the keys the
component renders, with no orphans.

Not run: the E2E suite, and the app itself.