fix(llm): batch attachment-path lookups for rule/corpus fingerprints (refs #680)

source.py: iter_rule_refs and iter_corpus_refs used to run one
_attachment_paths query per item just to build the cheap fingerprint,
before any change was even known. _attachment_paths_by_key replaces
that with one scan of the whole attachments table, used by
iter_rule_refs (only when at least one row lacks an anchor sha) and
iter_corpus_refs (every run). The per-item _attachment_paths calls on
the actual load path (_attachment_text, _rule_text,
_attachment_sections) are untouched — those only run for items that
already changed.

Also: the misplaced "rules" section marker now sits above the
rule-only helpers it should have covered from the start
(_attachment_text/_rule_text/rule_paragraphs/_anchor_doc, previously
stranded in the "comments" section), and iter_comment_refs's inner
_load no longer shadows the enclosing loop's row variable (renamed to
item_row).

src/llm/source.py

@@ -87,6 +87,26 @@ def _attachment_paths(store: Store, item_key: str) -> list[Path]:
     return [Path(r[0]) for r in rows]
 
 
+def _attachment_paths_by_key(store: Store) -> dict[str, list[Path]]:
+    """Every item's attachment paths, grouped by key, in one scan.
+
+    ``iter_rule_refs``/``iter_corpus_refs`` used to call
+    :func:`_attachment_paths` once per item just to build the
+    fingerprint — one query per rule/corpus item on every run, before
+    any change is even known. One table scan replaces that N+1."""
+    rows = (
+        store._con()
+        .execute(
+            "SELECT i.key, a.storage_path FROM attachments a JOIN items i ON i.id = a.item_id"
+        )
+        .fetchall()
+    )
+    by_key: dict[str, list[Path]] = {}
+    for key, path in rows:
+        by_key.setdefault(key, []).append(Path(path))
+    return by_key
+
+
 # ── comments ──
 
 
@@ -183,12 +203,12 @@ def iter_comment_refs(
                 _, body = parse_combined(combined.read_text())
                 if body.strip():
                     return Doc(key=key, text=body, metadata=meta, files=files)
-            row = (
+            item_row = (
                 store._con()
                 .execute("SELECT abstract FROM items WHERE key = ?", (key,))
                 .fetchone()
             )
-            abstract = (row["abstract"] if row else "") or ""
+            abstract = (item_row["abstract"] if item_row else "") or ""
             return (
                 Doc(key=key, text=abstract, metadata=meta) if abstract.strip() else None
             )
@@ -208,6 +228,9 @@ def iter_comment_docs(
             yield doc
 
 
+# ── rules ──
+
+
 def _attachment_text(store: Store, item_key: str) -> str:
     from rex.comments.combine import extract_attachment
 
@@ -265,9 +288,6 @@ def _anchor_doc(store: Store, item_key: str) -> tuple[str, str]:
     return (row[0], str(row[1])) if row else ("", "")
 
 
-# ── rules ──
-
-
 def iter_rule_refs(
     store: Store, *, keys: tuple[str, ...] = (), tag: str = ""
 ) -> Iterator[DocRef]:
@@ -291,6 +311,13 @@ def iter_rule_refs(
         )
         .fetchall()
     )
+    # Attachment paths are only needed for rows without an anchor sha;
+    # one batched scan replaces one _attachment_paths query per such row.
+    paths_by_key = (
+        _attachment_paths_by_key(store)
+        if any(not r["anchor_sha"] for r in rows)
+        else {}
+    )
     for row in rows:
         key = row["key"]
         if keys and key not in keys:
@@ -300,11 +327,7 @@ def iter_rule_refs(
             # item's own metadata (title, cms-rule: tag, date).
             fp = f"anchors:{row['anchor_sha']}|{row['updated_at']}"
         else:
-            fp = (
-                fingerprint_files(_attachment_paths(store, key))
-                + "|"
-                + row["updated_at"]
-            )
+            fp = fingerprint_files(paths_by_key.get(key, [])) + "|" + row["updated_at"]
 
         def _load(key=key) -> Doc | None:
             return _build_rule_doc(store, store.get(key))
@@ -524,11 +547,16 @@ def iter_corpus_refs(
         )
         + " ORDER BY i.id"
     )
-    for row in store._con().execute(sql, (tag,) if tag else ()).fetchall():
+    rows = store._con().execute(sql, (tag,) if tag else ()).fetchall()
+    # One batched scan replaces one _attachment_paths query per item —
+    # the fingerprint step used to cost a query per corpus item before
+    # any change was even known.
+    paths_by_key = _attachment_paths_by_key(store)
+    for row in rows:
         key = row["key"]
         if keys and key not in keys:
             continue
-        fp = row["updated_at"] + "|" + fingerprint_files(_attachment_paths(store, key))
+        fp = row["updated_at"] + "|" + fingerprint_files(paths_by_key.get(key, []))
 
         def _load(key=key) -> Doc | None:
             return _build_corpus_doc(store, store.get(key), zotero)