studio/frontend: cancelled message renders only action buttons when generation is stopped mid-thinking

[danielhanchen]

Repro

1. Load any reasoning-capable model (tested with unsloth/Qwen3.5-4B-MTP-GGUF, Think pill default off, so toggle it on).
2. Send a prompt that triggers a longer thinking phase, e.g. Write a long detailed essay (1500 words) on operating systems. Reply with the essay only, no <think>.
3. While the message still shows Generating... (before any visible answer tokens arrive), click the Stop button.

Observed

The assistant bubble is replaced by a floating row of 4 action buttons (Copy / Refresh / Delete / More) and nothing else. No message body, no "Generation stopped" indicator, no thinking-time chip. The composer is correctly re-enabled and the send button reappears, but the visible state suggests the message itself disappeared.

A user who is not looking at the network tab would reasonably ask: "did my stop click do anything?", "where is the partial answer?", or "is the chat broken?". This is the exact failure mode that produces the original "tools work, but I don't see anything happening or it looks stuck" complaint.

Why this matters

Studio's cancel path actually works (verified: /api/inference/load-progress resolves, llama-server stops streaming, the composer re-enables in well under 1s). The bug is purely visual: when there is no rendered content above the action-button row, the row floats with nothing to anchor it.

Suggested fix (frontend, smallest diff)

In studio/frontend/src/components/assistant-ui/thread.tsx (the AssistantMessage component, around the <MessagePrimitive.Root> near line 999 / 1166), gate the action-button row on the presence of message content OR show a placeholder when both content.length === 0 and status?.type === "incomplete" with reason: "cancelled". Two reasonable placeholders:

1. <span className="text-muted-foreground italic">Generation stopped</span> (short, terminal-state-aware)
2. Surface the partial reasoning panel even if its content was empty - Thought for {seconds}s reads better than blank.

Either is a one-component change. Both have precedent in the codebase: the thinking panel already collapses to "Thought for N seconds" once streaming stops, and the tool-fallback component already renders a "Cancelled tool" label when status.reason is "cancelled".

Screenshots

Pre-cancel: model is in thinking phase, stop button visible bottom right, only "Generating..." in middle of empty assistant area.

Post-cancel (i=05, ~3s after stop click): user prompt at top, 4 action buttons floating mid-screen, composer enabled at bottom. No message body. No "Generation stopped" label.

Environment

• Studio at unslothai/unsloth@a2f3793 plus PR studio: engage draft-mtp on vision MTP GGUFs (drop incorrect vision gate) #5560 (MTP-on-vision) applied
• Repro on unsloth/Qwen3.5-4B-MTP-GGUF and unsloth/Qwen3.5-9B-MTP-GGUF, UD-Q4_K_XL, 16384 context, f16 KV
• B200 GPU; happens within 1s of the stop click