[BUG] Cowork (web): near-identical assistant paragraphs re-rendered three times in one message block during/around auto-compaction

This concerns Cowork on claude.ai in the browser, not the CLI, so the CLI-only required fields are not applicable.

What happened

In a long Cowork session (cse_01W3ccWbTP8NWxio7Pu2qY5z), the transcript rendered three separate paragraphs that all say the same thing, stacked with no user turn between them:

Ep 37 is one edit away, blocked by a Google "Verify it's you" prompt on your client's account. Clear it here and I'll finish it.
Correction to what I said a minute ago: Ep 37 still carries it. ...

One thing from you unblocks it, a Google "Verify it's you" prompt sitting over Studio: clear it here, and Ep 37 takes a minute.
One video still has it. Google is asking to verify it's you before Studio will let me edit: clear that and it's done in a minute.

Each carries a different hyperlink to the same blocked step. To the user this reads as the assistant repeating itself compulsively. Auto-compaction ran immediately after ("Compacting our conversation so we can keep chatting… 84%").

Why this looks like a rendering/turn-boundary bug rather than model behaviour

• The paragraphs are not a single message's prose: they are three distinct final replies, each phrased as an opening line, concatenated into one visible block.
• No user message separates them in the rendered transcript, although the session did have user turns in between (a stop hook rejected two replies, each rejected attempt still reached the user).
• The session was at/near the compaction threshold when it happened.

Two things that would fix it

1. A reply that is refused by a Stop hook should not be rendered to the user as a delivered assistant message. Today a blocked reply is still shown, so the retry reads as repetition. (Related: there is no signal to the model that its previous attempt was already displayed.)
2. Assistant messages either side of an auto-compaction boundary should keep their turn separators in the rendered transcript instead of collapsing into one block.

Environment

• Cowork on claude.ai, Chrome on macOS
• Model: Opus 5
• Session: long-running, auto-compaction triggered

No repository files, tests, or entry points are named; start by reproducing the Cowork browser transcript behavior around Stop-hook rejections and auto-compaction. Done means refused replies are not shown as delivered messages and assistant turns on either side of compaction retain their separators.