fix(managers): address the E4 S3 code-review round 3 findings

- The tag menu's "trigger is gone" test now measures against the SCROLLER, not the
  window. The page scroller starts ~230px down (measured), so a row that slid under
  the fixed toolbar was still "on screen" by window coords and the menu floated over
  the chrome, anchored to something invisible.
- Scroll-repositioning keeps whichever side the menu is anchored to instead of
  resetting it to "below" every frame — that reset made the layout effect flip it
  back, costing two renders and two forced layouts per frame near the bottom edge.
- One placement rule (`coordsFrom`) now serves the initial open, the layout
  correction and every reposition, so those can no longer drift apart.
- Card page: reset on the filter/sort INPUTS only, and separately COMMIT the clamp
  when the page falls out of range. Keying the reset on the row count (round 2) was
  too blunt — a background refetch that merely added a channel threw the reader off
  page 7 for something they didn't do.
- Pager derives `pageCount()` once and `hasPagerContent` reuses it, instead of the
  same expression twice in one file.
- Re-indent the pagerNode JSX left dangling by round 2's ternary edit.

frontend/src/components/ChannelDiscovery.tsx

@@ -98,12 +98,13 @@ export default function ChannelDiscovery({
   const rows = (query.data ?? []).filter(
     (c) => !q || `${c.title ?? ""} ${c.handle ?? ""}`.toLowerCase().includes(q)
   );
-  // A re-sorted or RESIZED list invalidates the page you were on — reset to the first, as the table
-  // does on its own filter/sort changes. The row count is in the deps too, so a subscribe (which
-  // drops a channel from this tab) counts, not just an explicit filter change.
+  // Changing what you're looking at starts over at the first page, as the table does on its own
+  // filter/sort change. NOT keyed on the row count: subscribing drops a row from this tab, and being
+  // thrown to page 1 for it would punish the very action you just took. A list that shrank past your
+  // page is handled by committing the clamp instead (below).
   useEffect(() => {
     setCardPage(0);
-  }, [search, sort, rows.length]);
+  }, [search, sort]);
 
   const columns: Column<DiscoveredChannel>[] = [
     // Actions first, mirroring the Subscriptions table (where it sits right after Prio) so the two
@@ -202,6 +203,10 @@ export default function ChannelDiscovery({
   const sortedRows = isCards ? sortRows(rows, columns, sort) : rows;
   const cardTotalPages = cardSize <= 0 ? 1 : Math.max(1, Math.ceil(sortedRows.length / cardSize));
   const cardSafePage = Math.min(cardPage, cardTotalPages - 1);
+  // Commit the clamp when the list shrinks past your page, so it can't spring back later.
+  useEffect(() => {
+    if (cardPage > cardTotalPages - 1) setCardPage(cardTotalPages - 1);
+  }, [cardPage, cardTotalPages]);
   const pagedRows =
     !isCards || cardSize <= 0
       ? sortedRows

frontend/src/components/Channels.tsx

@@ -253,13 +253,14 @@ export default function Channels() {
     if (tagFilter.length && !tagFilter.some((id) => c.tag_ids.includes(id))) return false;
     return true;
   });
-  // A re-sorted or RESIZED list invalidates the page you were on, so go back to the first — the way
-  // DataTable resets its own page on a filter/sort change. Keyed on the resulting row count, not just
-  // the filter inputs, so an unsubscribe or a refetch that shrinks the list counts too: otherwise the
-  // clamp merely HIDES a stale page, which returns the moment the list grows again.
+  // Changing what you're looking at starts over at the first page, the way DataTable resets its own
+  // page on a filter/sort change. Deliberately NOT keyed on the row count: the ["channels"] query
+  // refetches on window focus and after every sync, so a catalogue that merely gained a channel would
+  // otherwise yank you off page 7 for something you didn't do. A list that shrank under you is handled
+  // by committing the clamp instead (see below).
   useEffect(() => {
     setCardPage(0);
-  }, [statusFilter, search, tagFilter, sort, visibleChannels.length]);
+  }, [statusFilter, search, tagFilter, sort]);
   const s = statusQuery.data;
 
   const onView = (c: ManagedChannel) =>
@@ -539,6 +540,12 @@ export default function Channels() {
   const sortedChannels = isCards ? sortRows(visibleChannels, columns, sort) : visibleChannels;
   const cardTotalPages = cardSize <= 0 ? 1 : Math.max(1, Math.ceil(sortedChannels.length / cardSize));
   const cardSafePage = Math.min(cardPage, cardTotalPages - 1);
+  // When the list shrinks under you the clamp above only HIDES the out-of-range page — commit it, or
+  // it springs back the moment the list grows again. Keeps your place otherwise (unlike a blanket
+  // reset), because it only fires when the page you're on no longer exists.
+  useEffect(() => {
+    if (cardPage > cardTotalPages - 1) setCardPage(cardTotalPages - 1);
+  }, [cardPage, cardTotalPages]);
   const pagedCards =
     !isCards || cardSize <= 0
       ? sortedChannels
@@ -852,54 +859,70 @@ function TagsCell({
     [fade.ref],
   );
 
-  // Anchor below the trigger; the layout effect below corrects it against the menu's REAL box. Only
-  // opens once a position exists, so the render gate can never show a stale anchor.
+  // ONE placement rule, so the position the menu opens at and the one it keeps while scrolling can't
+  // drift apart. `flip` keeps it anchored above the trigger once it has been flipped; the horizontal
+  // clamp uses the MEASURED width (`w-44` is rem — it grows with the text-size slider, so a px guess
+  // reads short at 1.3). Before first paint the menu isn't measurable yet; the layout effect below
+  // re-runs this once it is.
+  const coordsFrom = useCallback((r: DOMRect, flip: boolean) => {
+    const w = popRef.current?.offsetWidth ?? 0;
+    const left =
+      w && r.left + w > window.innerWidth - 8 ? Math.max(8, window.innerWidth - 8 - w) : r.left;
+    return flip
+      ? { left, bottom: window.innerHeight - r.top + 4 }
+      : { left, top: r.bottom + 4 };
+  }, []);
+
   const place = useCallback(() => {
     const r = btnRef.current?.getBoundingClientRect();
     if (!r) return false;
-    setCoords({ left: r.left, top: r.bottom + 4 });
+    setCoords(coordsFrom(r, false));
     return true;
-  }, []);
+  }, [coordsFrom]);
   const toggle = () => {
     if (open) setOpen(false);
     else if (place()) setOpen(true);
   };
 
-  // Correct the placement against the MEASURED box: flip above when it would run off the bottom,
-  // nudge back inside the right edge. Measured, not estimated — the menu's rows and its `w-44` are
-  // rem, so both grow with the text-size slider and any px guess reads short at 1.3.
+  // Correct the placement against the MEASURED box: flip above when it would run off the bottom (and
+  // there is room above), and apply the width clamp now that the menu can actually be measured.
   useLayoutEffect(() => {
     const pop = popRef.current;
     const r = btnRef.current?.getBoundingClientRect();
     if (!open || !pop || !coords || !r) return;
-    const { offsetHeight: h, offsetWidth: w } = pop;
-    const next = { ...coords };
-    // Flip up only when it actually overflows AND there is room above (else stay put and clip).
-    if (coords.top !== undefined && coords.top + h > window.innerHeight - 8 && r.top - h - 4 > 8) {
-      delete next.top;
-      next.bottom = window.innerHeight - r.top + 4;
-    }
-    if (coords.left + w > window.innerWidth - 8) next.left = Math.max(8, window.innerWidth - 8 - w);
+    const h = pop.offsetHeight; // the width is read inside coordsFrom, off the same node
+    const flip =
+      coords.top !== undefined
+        ? coords.top + h > window.innerHeight - 8 && r.top - h - 4 > 8 // overflows below, fits above
+        : true; // already flipped — keep it there
+    const next = coordsFrom(r, flip);
     if (next.left !== coords.left || next.top !== coords.top || next.bottom !== coords.bottom) {
       setCoords(next);
     }
-  }, [open, coords]);
+  }, [open, coords, coordsFrom]);
 
   // A fixed menu doesn't travel with its row, so FOLLOW the trigger on scroll rather than closing:
   // in the table the menu used to scroll with the row, and closing on every wheel tick would make
-  // attaching several tags a reopen-per-tag chore. Closes only once the trigger leaves the viewport
-  // (nothing left to anchor to) or on resize. Scrolling INSIDE the menu must not move it.
+  // attaching several tags a reopen-per-tag chore. Closes only once the trigger leaves the SCROLLER
+  // (not the window — the page scroller starts below ~230px of fixed chrome, so a row hidden behind
+  // the toolbar is still "on screen" by window coords and the menu would float over the chrome).
+  // Scrolling INSIDE the menu must not move it.
   useEffect(() => {
     if (!open) return;
     let raf = 0;
     const reposition = () => {
       raf = 0;
       const r = btnRef.current?.getBoundingClientRect();
-      if (!r || r.bottom < 0 || r.top > window.innerHeight) {
+      const scroller = document.querySelector("main")?.getBoundingClientRect();
+      const topEdge = scroller ? scroller.top : 0;
+      const bottomEdge = scroller ? scroller.bottom : window.innerHeight;
+      if (!r || r.bottom < topEdge || r.top > bottomEdge) {
         setOpen(false);
         return;
       }
-      setCoords({ left: r.left, top: r.bottom + 4 });
+      // Keep whichever side it is anchored to, so the layout effect has nothing left to correct —
+      // resetting to "below" every frame would make it flip back and forth, two renders per frame.
+      setCoords((cur) => (cur ? coordsFrom(r, cur.bottom !== undefined) : cur));
     };
     const onScroll = (e: Event) => {
       if (popRef.current && e.target instanceof Node && popRef.current.contains(e.target)) return;
@@ -913,7 +936,7 @@ function TagsCell({
       window.removeEventListener("scroll", onScroll, true);
       window.removeEventListener("resize", onResize);
     };
-  }, [open]);
+  }, [open, coordsFrom]);
   if (userTags.length === 0) return null;
   // Show only the tags actually attached to this channel; the “+” opens a per-channel
   // picker to attach/detach (kept the convenient "kirakás" without listing every tag

frontend/src/components/Pager.tsx

@@ -7,12 +7,16 @@ import { ChevronLeft, ChevronRight } from "lucide-react";
 // size and slices its own rows; this only renders the controls and reports changes. `sizeLabel` lets
 // the caller say "Rows per page" or "Cards per page". Returns null when there's nothing to page.
 
+/** How many pages `total` rows make at `size` (`size <= 0` = "All" = one page). */
+function pageCount(total: number, size: number): number {
+  return size <= 0 ? 1 : Math.max(1, Math.ceil(total / size));
+}
+
 /** Whether a pager would render anything — the ONE owner of that rule. Callers that wrap the pager
  * in their own chrome gate on this instead of re-deriving it: a drifted second copy is exactly what
  * left empty control bands behind on tables with nothing to page. */
 export function hasPagerContent(total: number, size: number): boolean {
-  const totalPages = size <= 0 ? 1 : Math.max(1, Math.ceil(total / size));
-  return totalPages > 1 || total > 0;
+  return pageCount(total, size) > 1 || total > 0;
 }
 export default function Pager({
   page,
@@ -35,8 +39,7 @@ export default function Pager({
   sizeLabel: string;
 }) {
   const { t } = useTranslation();
-  const allRows = size <= 0;
-  const totalPages = allRows ? 1 : Math.max(1, Math.ceil(total / size));
+  const totalPages = pageCount(total, size);
   const safePage = Math.min(page, totalPages - 1);
 
   // Local text for the editable "jump to page" box (committed on Enter/blur).