Fix stale pinned-brief images; enrich all 7 + retry failures

Root cause (Codex audit): the client pins the brief by generated_at, but image
enrichment populates image_url AFTER the brief is built without bumping
generated_at — so a verbatim pinned copy stays imageless even once the server
has the image. The reclassify rebuilt the brief and the early pin stuck.

- Frontend: when reusing a pinned brief (same generated_at), refresh server-owned
  metadata by article id (esp. image_url) while preserving the user's order and
  replacements. Re-saves the merged view so it stays current.
- enrich_brief_images: default limit 5 -> 7 (any brief item can become the hero
  via the client fallback or a replace, so cover the whole brief).
- Don't cache image failures forever: retry brief items still missing an image
  after a TTL (retry_days=2) instead of stamping them imageless permanently.

Pairs with the hero image fallback (dd0087b). 99 tests pass.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

goodnews/enrich.py

@@ -146,27 +146,36 @@ def fetch_og_image(url: str | None) -> str | None:
     return None  # too many redirects
 
 
-def enrich_brief_images(conn: sqlite3.Connection, brief_date: str, fetch=fetch_og_image, limit: int = 5) -> int:
-    """Fetch a hero-quality image for brief items that lack one (once each).
+def enrich_brief_images(
+    conn: sqlite3.Connection, brief_date: str, fetch=fetch_og_image, limit: int = 7, retry_days: int = 2
+) -> int:
+    """Fetch a hero-quality image for brief items that lack one.
 
-    Only touches the given date's brief items with no image and not yet checked;
-    stamps image_checked_at either way so failures are not retried forever.
+    Any of the brief's items can become the hero (via the client's fallback or a
+    replace), so this covers the whole brief (limit defaults to the brief size, 7),
+    not just the top few. Items already carrying an image are left alone; items
+    still without one are retried after `retry_days` so a transient fetch failure
+    or a weaker earlier extractor doesn't mark an article imageless forever.
     Returns how many images were newly found.
     """
-    # Brief items not yet checked. We fetch even when a feed image exists,
-    # because feed thumbnails are often tiny and the hero is shown large — a
-    # page's og:image (the social-share image) is the better hero visual.
+    # Fetch even when a feed image exists, because feed thumbnails are often tiny
+    # and the hero is shown large — a page's og:image is the better hero visual.
     rows = conn.execute(
         """
         SELECT a.id, a.canonical_url
         FROM daily_briefs b
         JOIN daily_brief_items bi ON bi.brief_id = b.id
         JOIN articles a ON a.id = bi.article_id
-        WHERE b.brief_date = ? AND a.image_checked_at IS NULL
+        WHERE b.brief_date = ?
+          AND (
+            a.image_checked_at IS NULL
+            OR ((a.image_url IS NULL OR a.image_url = '')
+                AND a.image_checked_at < datetime('now', ?))
+          )
         ORDER BY bi.rank
         LIMIT ?
         """,
-        (brief_date, limit),
+        (brief_date, f"-{retry_days} days", limit),
     ).fetchall()
 
     found = 0