search: email address search + address-first ranking; inline images; folder i18n; link security

Search (PG backend):
- '@' terms match address substrings (from/to/cc), so a partial email like
  "dyrektor@wfdif" resolves and narrows predictably as you type
- index each address split into parts, so a bare domain/local word ("wfdif")
  finds the address (indexed as one lexeme otherwise); dotted terms fall back to
  an address substring
- rank sender/recipient name+address matches above body/subject-only ones
  (addr_hit tier, then recency); cursor extended to keep keyset pagination correct
- scripts/reindex_address_parts.py backfills address tokens into existing rows

Inline images:
- extract attachments via a full MIME walk instead of shallow iter_attachments(),
  which missed inline images nested in mixed>related>alternative (cid: rendered
  broken); has_attachments counts only real (non-cid) attachments
- scripts/backfill_inline_attachments.py stores previously-missed inline images

Folder-name i18n (SPEC §12 M2):
- accent/punctuation-insensitive, multi-delimiter (/ \ .) special-use inference
  across ~13 languages; curated names (no substring false positives)

Link security (reading panel):
- links never navigate the message iframe: a parent-side click interceptor opens
  http/https in a new tab (noopener) and mailto in the composer (recipient
  pre-filled); all other schemes blocked

Backend 292 tests green, ruff + mypy --strict clean; frontend tsc/eslint/vitest clean.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01U72FZiF1W1pb8XxyWBTREA

backend/app/indexer/indexer.py

@@ -54,7 +54,7 @@ def _build_search_doc(message: Message, parsed: ParsedEmail, special_use: str |
         cc=parsed.cc_addr,
         subject=parsed.subject,
         body_text=parsed.text_body,
-        attachment_names=[a.filename for a in parsed.attachments],
+        attachment_names=[a.filename for a in parsed.attachments if a.content_id is None],
         has_attachments=parsed.has_attachments,
         is_unread="\\Seen" not in message.flags,
         is_starred="\\Flagged" in message.flags,

backend/app/indexer/mime.py

@@ -46,7 +46,9 @@ class ParsedEmail:
 
     @property
     def has_attachments(self) -> bool:
-        return len(self.attachments) > 0
+        # Inline resources (cid: images) carry a content_id and render in-body, not
+        # as attachment chips — only real attachments flip the paperclip.
+        return any(a.content_id is None for a in self.attachments)
 
     @property
     def snippet(self) -> str:
@@ -85,17 +87,25 @@ def _get_body_content(msg: EmailMessage, subtype: str) -> str | None:
 
 
 def _extract_attachments(msg: EmailMessage) -> list[ParsedAttachment]:
+    # Walk the whole tree, not iter_attachments(): the latter is shallow and treats
+    # multipart/related as a body, so an inline image nested in mixed>related>alt is
+    # never yielded (its cid: reference then renders broken). A part is an attachment
+    # or inline resource if it has a filename, a Content-ID, or attachment disposition;
+    # the body text/plain and text/html parts have none of these and are skipped.
     out: list[ParsedAttachment] = []
-    for part in msg.iter_attachments():
-        if not isinstance(part, EmailMessage):
+    for part in msg.walk():
+        if part.is_multipart() or not isinstance(part, EmailMessage):
+            continue
+        cid = part.get("Content-ID")
+        filename = part.get_filename()
+        if not (filename or cid or part.get_content_disposition() == "attachment"):
             continue
         payload = part.get_payload(decode=True)
         if not isinstance(payload, bytes):
             continue
-        cid = part.get("Content-ID")
         out.append(
             ParsedAttachment(
-                filename=part.get_filename() or "attachment",
+                filename=filename or "attachment",
                 mime_type=part.get_content_type(),
                 content_id=cid.strip("<>") if cid else None,
                 payload=payload,