fix(agent-chat): stale "subagents running" indicators — settle non-terminal subagents in the transcript layer

[Zeus-Deus]

Fixes #153

Problem

The frontend transcript layer had no settlement path for non-terminal subagents: the only running→terminal transition was a terminal subagent_updated snapshot. When the Claude adapter's demux loses track (sidecar restart/resume), the spawning tool's raw parent-scoped tool_result flows through and no terminal snapshot ever arrives — so the SubagentActivityBar / SubagentsCard showed "N subagents running" forever, the thread persisted that way, and reopening the workspace resurrected running spinners on a dead session. Provider rate-limit stops were invisible (runtime_warning → console.warn only).

Fix (frontend-only — the backend SubagentTracker already handles the sidebar dot)

1. tool_result derivation — a parent-scoped tool_result whose tool_use_id matches a running subagent's demux id or its parentItemId (spawning call id, now merged from SubagentSnapshot.parent_item_id) settles it (completed/failed). The Workflow tool's own tool_result likewise settles the run + its in-flight phase agents. Because tool_results are persisted, existing stuck threads self-heal on hydrate replay — no migration.
2. Forced settle — session_state_changed{closed|error} and a new user turn (mirror of SubagentTracker::clear_thread; queued follow-ups behind an active turn are left alone) interrupt running subagents and stop running workflows, even when streaming was already false.
3. Interrupted-on-hydrate — new view-only interrupted status (SubagentViewStatus, never on the wire): muted-amber non-spinning glyph, excluded from runningSubagentEntries. Post-replay reconciliation force-settles anything still running — a dead session never hydrates a spinner. A pending_approval workflow is preserved (a legitimate resting state, not a spinner).
4. Revivable assumed settles — every inferred settlement stamps statusAssumed; a later real running snapshot (Claude background task_progress ticks) revives the row, a later real terminal snapshot confirms and clears the flag. Self-healing on live streams and on replay.
5. runtime_warning visibility — a conservative classifier (runtime-notice.ts) promotes user-facing warnings (rate-limit status: "rejected", enumerated assistant error: …) to an inline amber runtime_notice transcript row; SDK debug noise stays console-only.
6. Hardening found in review — a status-less snapshot (Rust serde-defaults status to pending; hand-built payloads can omit it) leaked undefined into the view and crashed SubagentRow at render. mergeStatus now treats missing/unknown incoming status as pending.

Verification

• npm run check clean, npm run test 2415/2415 (new: settlement/revive/classifier/hydrate-reconciliation coverage; dev-fixture tests deliberately updated — the hydrated seed's still-running subagent now correctly settles to interrupted).
• Visually verified in the dev mock (browser pane): the hydrated seed shows the muted-amber ⊘ interrupted row with no stale docked bar; a live streamSubagents() run still shows the running bar, then settles green end-to-end.
• Rust untouched (should_persist_event already persists everything needed).