Surface non-standard LLM finish reasons (max_tokens, content_filter, etc.) as UI notifications, not just console warnings

Problem

When a model stops for a reason other than a clean completion — hitting max_tokens, exceeding the context window, getting filtered by content moderation — ellmer and chatlas already detect this (AssistantTurn.finish_reason / check_finish_reason()) and warn about it. But that warning is console-only. In a Shiny Chat app, it's invisible: the user just sees whatever failure happens downstream (e.g. a confusing JSON parse error on truncated output), with no clue that the actual cause was a token limit.

This came up concretely in querychat: a truncated response caused a parse error: premature EOF in the chat UI, while the real cause — "Response was truncated because it hit the max_tokens limit" — only ever printed to the R console.

Proposed direction

• ellmer and chatlas each upgrade their existing plain warning (cli::cli_warn() / warnings.warn(..., UserWarning)) to a classed/typed signal that carries the finish reason as data — an ellmer_finish_reason rlang condition class in R, an IncompleteResponseWarning subclass in Python (chatlas already has this exact pattern for a different case, ToolFailureWarning).
• shinychat's R (chat_append_stream()) and Python (append_message_stream()) sides each catch that signal while consuming the stream and surface it as a UI notification (showNotification() / notification_show()), matching the type="warning" treatment each side already uses elsewhere for errors.
• No public API signature changes anticipated on either side.

A full design (data flow, exact call sites, edge cases, one open technical assumption still needing a smoke test) is written up and available on request — happy to paste it in or link it once it has a stable home.

Scope

This covers ellmer, chatlas, and shinychat (R + Python). It does not cover querychat's own downstream JSON-parsing error handling, which is a separate, narrower fix.

Read the existing finish-reason handling around AssistantTurn.finish_reason and check_finish_reason(), then inspect shinychat's chat_append_stream() and append_message_stream() consumers. Verify the proposed signal paths and the open smoke-test assumption; done means non-standard finish reasons appear as warning notifications in both R and Python streaming UIs without public API signature changes.