fix(web): isSending honors activeTurnId so thinking indicator survives stale terminal events

[vellum-apollo-bot]

When a stale message_complete (or assistant_activity_state{phase:idle}) from a previous turn lands during the await sendMessageViaStream(...) window inside sendMessage, completeTurn() resets phase: idle and clears activeTurnId. The subsequent acceptSend(turnId) restores activeTurnId but does not touch phase or lastTerminalReason, leaving the turn store in an illegal but observable shape:

phase: "idle"
activeTurnId: "turn-NEW"
lastTerminalReason: "complete"

The old phase-only isSending predicate returned false here, so shouldShowThinkingIndicator failed on its first AND-clause and the thinking indicator stayed hidden until the first assistant delta arrived. _vellumDebug.chat.thinkingIndicator() reported failingConditions: ["notSendingAndNotRestoredProcessing"] in exactly this shape in the wild.

Fix

Treat activeTurnId !== null as authoritative for in-flight state. Keep the existing phase checks so the post-complete "queued waiting" state (phase: "queued", activeTurnId: null, pendingQueuedCount > 0) still reports as sending.

export function isSending(state: TurnState): boolean {
  return (
    state.activeTurnId !== null ||
    state.phase === "queued" ||
    state.phase === "thinking" ||
    state.phase === "streaming" ||
    state.phase === "awaiting_user_input"
  );
}

Once the first assistant delta arrives, onTextDelta transitions phase → "streaming". The indicator continues to render through the rest of the turn from the streaming/thinking phases as before.

Why not fix the stale-terminal race itself?

The deeper fix is to gate handleMessageComplete / handleAssistantActivityState on a turn-id match (the daemon already carries requestId on activity-state events; message-complete would need to grow one). That requires daemon coordination and is a separate concern. This PR fixes the symptom at the predicate that all consumers read from.

Tests

• turn-store.test.ts — three new cases pinning the predicate to the actual invariant: returns true for the illegal shape, false for terminal idle, true for queued-waiting.
• use-message-reconciliation.test.tsx — updated does NOT call onPollReconciled when turn is already idle to match its name (genuine terminal idle has activeTurnId: null). The illegal-state recovery is already covered by the dedicated "stuck sending" rescue tests above it.
• 268 tests pass across the touched areas.

[dvargasfuertes]

Check this out. Based on the comment, we expected the state.phase to be thinking, instead of adding this active turn id null check, which I think would always render the thinking indicator during the lifetime of the assistant response

[vellum-apollo-bot]

Good catch — you're right. The fix belongs in the state machine, not the predicate.

Reverted isSending to phase-only. Moved the recovery into acceptSend (and the matching USER_SEND_ACCEPTED reducer case): when the POST resolves and we're entering the "waiting for first delta" window, restore phase: "thinking" + lastTerminalReason: null if a stale terminal landed between requestSend and acceptSend. Non-terminal phases (e.g. awaiting_user_input from a surface that's still active) are left alone.

This matches the isThinking doc comment exactly — phase IS thinking through that window, so all downstream selectors keep working unchanged.

Pushed as 24c7b2f.