The picker's paging state outlives a reopen — navigate, don't assume

Sleeping Gods and Gloomhaven "vanished on the second pass" because the
second pass began wherever the first ended: closing and reopening the
version sub-view does NOT reset it to page 1 (Angular keeps the scope),
so the rescan started mid-list and never revisited the earlier pages
holding the row. Verified against the live picker.

The second pass now clicks the visible numbered "1" anchor first — and
so does the initial scan, since paging state can outlive anything. The
reopen is gone entirely. Docs record both this and the has_text
whitespace trap.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016jXZFSTZQKzAC8fqpWSz9g

data/upload_log.csv

@@ -15,3 +15,8 @@ add,419687,,Munchkin Big Box,711400,,added,2026-08-06T02:45:16+00:00,
 add,252153,,Tang Garden,404702,,added,2026-08-06T02:45:23+00:00,
 add,255984,,Sleeping Gods,701034,,failed,2026-08-06T02:45:58+00:00,"TimeoutError: Locator.click: Timeout 30000ms exceeded. Call log: - waiting for get_by_role(""dialog"").locator(""li:has(.summary-item-thumbnail)"").filter(has_text=re.compile(r""Sleeping\ Gods\ \(English\ Gamefound\ edition\)\ \(2023\)"", re.IGNORECASE)).first"
 add,283766,,Sleeping Gods: Tides of Ruin,463140,,added,2026-08-06T02:46:07+00:00,
+add,255984,,Sleeping Gods,701034,,failed,2026-08-06T02:49:44+00:00,RuntimeError: 'Sleeping Gods (English Gamefound edition) (2023)' was listed a moment ago but vanished on the second pass — retryable
+add,280794,,Etherfields,458300,,added,2026-08-06T02:49:53+00:00,
+add,380837,,Botany,650545,,added,2026-08-06T02:50:00+00:00,
+add,393672,,Gloomhaven: Buttons & Bugs,669782,,failed,2026-08-06T02:50:07+00:00,RuntimeError: 'Gloomhaven: Buttons & Bugs (English edition) (2024)' was listed a moment ago but vanished on the second pass — retryable
+add,14535,,SPANC: Space Pirate Amazon Ninja Catgirls,29075,,added,2026-08-06T02:50:15+00:00,

docs/bgg-upload-flow.md

@@ -128,9 +128,15 @@ earlier walkthrough had marked *verified*, and the parts marked
   inside `<li class="visible-xs-*">`. A selector matches both, and
   `.first` may be the hidden one — Playwright then waits for it to become
   visible until it times out. Always click the first *visible* match.
-  First/Prev exist ONLY in the mobile set, so they are unclickable on a
-  desktop viewport: to return to page 1, close and reopen the sub-view
-  (it always opens on page 1) and step forward with Next.
+  First/Prev may be unclickable on a desktop viewport (mobile-only
+  variant). To return to page 1, click the visible numbered **"1"**
+  anchor: the sub-view's paging state SURVIVES closing and reopening it
+  (Angular keeps the scope), so a reopen lands wherever it was left, not
+  on page 1.
+- **Match row text in Python, not with `has_text`.** Playwright's
+  `has_text` regex tests raw `textContent`, which carries the markup's
+  tabs and newlines; a whitespace-normalized capture will never equal it.
+  Normalize both sides yourself and click the row by index.
 - **Row text is `<game name> (<version name>) (<year>)`**, and the game
   name is localized (a Czech edition's row starts "Spící bohové"). Match
   the version name inside its parentheses.

src/bggpipe/upload.py

@@ -382,6 +382,7 @@ class PlaywrightUploader:
     # visible-xs-* — so every one of these selectors matches hidden nodes
     # too; clicking one waits forever. Always pick the visible match.
     _NEXT_PAGE = 'ul.pagination a[title="Next Page"]'
+    _PAGE_LINKS = "ul.pagination a"
     # Rows read "<game name> (<version name>) (<year>)", so the version name
     # is matched inside its parentheses — bare substrings would let
     # "English edition" match "English edition, second printing".
@@ -428,6 +429,8 @@ class PlaywrightUploader:
         # BGG's API name can carry a printing qualifier the picker omits
         # ("English edition 2018-2" vs "(English edition) (2018)")
         relaxed = f"({self._TRAILING_YEAR.sub('', version_name).strip()})".casefold()
+        self._goto_first_page(dialog)  # a fresh sub-view opens on page 1,
+        # but never assume it: paging state outlives a close/reopen
         exact: list[tuple[int, str]] = []
         loose: list[tuple[int, str]] = []
         pages = 0
@@ -472,18 +475,28 @@ class PlaywrightUploader:
             f"({', '.join(t for _, t in hits[:3])}…) — refusing to guess"
         )
 
-    def _click_row(self, dialog, text: str) -> bool:
-        """Reopen the sub-view (it always starts on page 1) and hunt for the
-        row again by its NORMALIZED text, clicking by index.
+    def _goto_first_page(self, dialog) -> None:
+        """Back to page 1 via the numbered "1" anchor. Closing and
+        reopening the sub-view does NOT reset Angular's paging state (it
+        reopens wherever it was left), and First/Prev render in a
+        mobile-only variant that a desktop viewport can never click."""
+        pager = dialog.locator(self._PAGE_LINKS)
+        for i in range(pager.count()):
+            anchor = pager.nth(i)
+            if anchor.is_visible() and (
+                " ".join((anchor.text_content() or "").split()) == "1"
+            ):
+                anchor.click()
+                self._page.wait_for_timeout(400)
+                return
+        # no pagination rendered: a single-page list is already page 1
 
-        Two traps this avoids: First/Prev exist only in the mobile variant
-        (unclickable on a desktop viewport), and Playwright's has_text
-        regex matches raw textContent — whose tabs and newlines a
-        whitespace-normalized capture will never equal."""
-        dialog.get_by_role("button", name="Cancel").first.click()
-        self._page.wait_for_timeout(300)
-        dialog.get_by_role("button", name="Set version/edition").click()
-        dialog.locator(self._VERSION_ROWS).first.wait_for(timeout=15_000)
+    def _click_row(self, dialog, text: str) -> bool:
+        """Find the row again by its NORMALIZED text and click it by index.
+        Playwright's has_text regex matches raw textContent — whose tabs and
+        newlines a normalized capture will never equal — so the comparison
+        happens in Python, and the click addresses a position."""
+        self._goto_first_page(dialog)
         for _ in range(MAX_VERSION_PAGES):
             for i, row_text in enumerate(self._rows_on_page(dialog)):
                 if row_text == text:

tests/test_upload.py

@@ -551,8 +551,9 @@ def test_undeterminable_state_raises_instead_of_guessing(monkeypatch):
 
 
 class _FakePicker:
-    """Minimal stand-in for the picker sub-view: pages of row texts, a
-    Next anchor that disables on the last page."""
+    """Stand-in for the picker sub-view. Models the two traps the real one
+    sets: paging controls render twice (a hidden mobile duplicate first),
+    and paging state SURVIVES closing and reopening the sub-view."""
 
     def __init__(self, pages):
         self.pages = pages
@@ -562,9 +563,12 @@ class _FakePicker:
         self.cancelled = False
         self.reopened = 0
 
-    # -- locator plumbing the uploader uses ---------------------------
     def locator(self, selector):
         picker = self
+        is_next = 'title="Next Page"' in selector
+        is_pager = "pagination" in selector
+        # anchors: [hidden mobile "1", "1", "2", ...] mirroring the real DOM
+        anchors = ["1"] + [str(n + 1) for n in range(len(picker.pages))]
 
         class _Loc:
             def __init__(self, index=0):
@@ -574,22 +578,25 @@ class _FakePicker:
                 return picker.pages[picker.page]
 
             def count(self):
-                if "pagination" in selector:
-                    return 2  # BGG renders a mobile AND a desktop control
+                if is_next:
+                    return 2  # hidden mobile + visible desktop
+                if is_pager:
+                    return len(anchors)
                 return len(picker.pages[picker.page])
 
             def nth(self, i):
                 return _Loc(i)
 
-            def is_visible(self):
-                # among PAGING controls, index 0 is the mobile
-                # (visible-xs-*) duplicate — present but not clickable
-                return "pagination" not in selector or self.index != 0
-
             @property
             def first(self):
                 return _Loc(0)
 
+            def is_visible(self):
+                return not (is_pager and self.index == 0)
+
+            def text_content(self):
+                return anchors[self.index] if is_pager and not is_next else ""
+
             def wait_for(self, **kw):
                 return None
 
@@ -598,16 +605,17 @@ class _FakePicker:
                 return "ng-scope disabled" if last else "ng-scope"
 
             def click(self):
-                if "pagination" in selector and not self.is_visible():
+                if is_pager and not self.is_visible():
                     raise AssertionError("clicked a HIDDEN paging control")
-                if "Next Page" in selector:
+                if is_next:
                     picker.page += 1
-                elif "summary-item-thumbnail" in selector:
+                elif is_pager:
+                    picker.page = int(anchors[self.index]) - 1
+                else:
                     picker.clicked_index = self.index
                     picker.clicked = picker.pages[picker.page][self.index]
 
             def filter(self, has_text=None):
-                picker.clicked = has_text.pattern if has_text else None
                 return self
 
         return _Loc()
@@ -624,8 +632,7 @@ class _FakePicker:
                 if name == "Cancel":
                     picker.cancelled = True
                 elif name == "Set version/edition":
-                    picker.page = 0  # the sub-view reopens on page 1
-                    picker.reopened += 1
+                    picker.reopened += 1  # note: does NOT reset the page
 
         return _Btn()
 
@@ -697,8 +704,7 @@ def test_paging_never_clicks_a_hidden_mobile_control():
     )
     picked, _ = up._select_version(dialog, "English edition 2009")
     assert picked is True  # no AssertionError from the fake = no hidden click
-    # the initial open plus ONE reopen — never a "First Page" click
-    assert dialog.reopened == 2
+    assert dialog.clicked_index == 0  # found on the later page, clicked
 
 
 def test_outstanding_failures_ignores_retried_jobs():
@@ -731,3 +737,20 @@ def test_outstanding_failures_ignores_retried_jobs():
     ]
     assert outstanding_failures(rows) == 1  # game 1 was retried and landed
     assert outstanding_failures([]) == 0
+
+
+def test_paging_state_survives_reopen_so_page_one_is_clicked():
+    """The real sub-view reopens wherever it was left, so the second pass
+    must navigate back to page 1 explicitly — Sleeping Gods and Gloomhaven
+    both "vanished" when it didn't."""
+    up, dialog = _picker_uploader(
+        [
+            ["Sleeping Gods (English Gamefound edition) (2023)"],
+            ["Sleeping Gods (German edition) (2022)"],
+        ]
+    )
+    dialog.page = 1  # left on the last page by a previous scan
+    picked, why = up._select_version(dialog, "English Gamefound edition")
+    assert picked is True
+    assert why == ""
+    assert dialog.clicked.startswith("Sleeping Gods (English Gamefound")