Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In
`@apps/web/app/`(app)/workspaces/[workspaceId]/unify/feedback-records/components/feedback-records-table.tsx:
- Around line 141-146: The toast error messages are hiding server-action
details; update both failure branches that handle the result from fetchRecords
(the one after const result = await fetchRecords("refresh") and the later branch
that currently calls toast.error(t(...))) to include the server error text by
calling toast.error(getFormattedErrorMessage(result), { id: toastId }) (or
toast.error(getFormattedErrorMessage(result)) if no id is used), ensuring you
import/consume the existing getFormattedErrorMessage helper and preserve the
localized fallback when result has no server message.
- Around line 159-176: handleRefresh and handleLoadMore can run concurrently and
the last finisher overwrites records/cursors; add a single mutual-exclusion
guard to prevent overlapping fetches by introducing and checking a shared lock
(e.g., isFetchingRef or fetchMutex) at the start of both handlers (handleRefresh
and handleLoadMore), bail out if already locked, then set the lock before
calling fetchRecords and clear it in a finally block; ensure you still set the
per-action UI flags (isLoadingMore/isRefreshing) but only update setRecords and
setCursors when the lock is held (or by the winning operation) so pagination
state cannot be clobbered.

In `@apps/web/app/`(app)/workspaces/[workspaceId]/unify/feedback-records/page.tsx:
- Around line 47-54: The bug: initialCursors is built from each FRD's full
first-page next_cursor but the rendered dataset is globally sliced to 50, so
some FRD first-page items may be omitted and later become unreachable. Fix by
computing initialCursors only for FRDs whose first-page items are included in
the initial displayed set (i.e., derive the map from the dataset after the
global slice), or by building the cursor map from the full per-FRD results
before slicing and then tracking which FRDs contributed items to the rendered 50
so you only advance cursors for those FRDs; update the logic around
initialCursors, frds and results (or the pre-slice result variable) to reflect
this consistent approach so load-more uses cursors only for FRDs that actually
contributed to the initial render.