Opt-in hover strip for the model each assistant reply used

[thehiddenlotus]

After a thread uses more than one model, the composer footer only shows what will send next. Old assistant replies do not say which model, Fast, or Effort they ran with, or what that reply cost. Mixed-provider threads become hard to audit.

#2601 tried to solve a related need with an always-visible stats footer plus persist, tok/sec, and TTFT. That was an XL and got closed. I do not want that PR. I want the small part: provenance on the existing hover row.

Proposed scope, one problem:

• Device-local Appearance checkbox, default off
• When on, the existing assistant hover row (Copy / Branch / time) gets a muted strip on the right: Auto, catalog label, Fast, Effort, compact #
• Catalog label is the same name the model picker already shows. Fast and Effort mean what the composer already means. # is stored parent uncached input plus output, omitted when missing
• UI on data the chat already has. No persist, no routed-model lookup, no tok/sec, no new footer, no live-stream chrome

Out of this request: writing usage at persist time, resolving Auto to the routed model, provider color on #, showing the hover row while a reply is still streaming.

I have a local implementation of that slice and can open a PR if this scope is acceptable.

Please reply with an explicit yes on this direction and scope if you want it. CONTRIBUTING asks for that comment before a feature PR.