fix(site/src/pages/AgentsPage): clear streamed message when a chat errors - #28277

…rors

When a chat stopped with an error, the WebSocket error handler set the
chat status and stored the error but never cleared the in-flight stream
state. A tool call that had no result yet kept its running status, so
the partial output (for example "Creating workspace...") kept showing a
spinner even though the chat had errored. Refreshing showed the correct
status because the stream state is transient.

Treat the error event as a terminal turn boundary. Discard buffered
parts, gate late parts from the finished turn, and clear the stream
state. The error callout still surfaces the failure.

…r clearing stream

Add a ChatPageContent story that streams a create_workspace tool call and
applies the terminal error, asserting the running row is removed and the
failure callout stays. Also drop test comments that restated the
assertions below them.

… story

Mount the running tool state, then trigger the terminal error during play
and assert the running row is removed while the failure callout remains.
Previously the story set the cleared state before mount, so it passed even
if the error path stopped clearing the stream.

A prior commit set streamReportedWaiting on the error path so late parts
from the errored turn would be dropped. The gate is only reopened by a
status event, and the server delivers parts and status on separate
unordered paths, so a new turn's first part can arrive before its running
status and be silently dropped, corrupting the response.

Drop the gate and keep clearing the stream state and buffered parts.
This preserves the fix for the leftover spinning tool while never
dropping a part that belongs to the next turn. A part already queued
from the errored turn can still briefly reappear afterward, which is a
cosmetic race that the next status event or a refresh resolves.

…at error

A part already queued from an errored turn can arrive after the error and
repopulate the cleared stream. Render the failed state without that
leftover so it does not show as a live row, and clear the stored error
when a new turn's status arrives so the next stream renders. Also reset
the story store between runs so the error transition is exercised each
time.

…inal errors

The render suppression forced hasAccumulatedOutput off whenever streamError
was set, but request and parse errors set streamError while the stream is
intentionally still running. Gate the suppression on the chat status being
error, which only the terminal stream error sets, so a genuinely live
assistant row keeps rendering for non-terminal errors.

…to terminal status

Clear the stored stream error only when an errored chat begins a new
turn, so a status from a still-running turn does not wipe an unrelated
request-failure callout. Suppress stream output whenever the chat status
is error, not only while the error is set, so a stale leftover does not
reappear as a streaming row while a retry is in flight.

…connect phases

The retry and reconnect branches still passed the unsuppressed
accumulated-output flag, so a stale leftover from a failed turn could
render under a reconnect banner while the chat status is error. Pass the
terminally suppressed value through those branches too, and tighten the
comment to state only the non-obvious invariant.