fix(search): pure BM25 relevance ranking + camelCase token splitting

Two search quality fixes:

1. Remove fan_in popularity boost from BM25 ranking.
   The fan_in>5 clause added +3.0 to the rank of heavily-called functions,
   causing popular-but-irrelevant results (e.g. 'update' with fan_in=222) to
   outrank relevant matches. search_graph with query= is now pure BM25
   relevance + label-type differentiation only. in_deg/out_deg are still
   returned in results for display but do not affect sort order.

2. Add cbm_camel_split() SQLite function for FTS5 indexing.
   FTS5's unicode61 tokenizer treats 'createSession' as a single token
   'createsession'. Searching for 'session' would not match it.
   cbm_camel_split() expands camelCase names into space-separated tokens:
   'createSession' → 'createSession create Session'
   'HTMLParser' → 'HTMLParser HTML Parser'
   The original name is preserved as the first token for exact-match queries.

   Applied in: FTS5 triggers (INSERT/DELETE/UPDATE on nodes table) and
   the bulk FTS5 backfill after full indexing.

3. Switch FTS5 from content='nodes' (external content) to content=''
   (contentless). External content mode re-verifies matches against the
   source table at query time, which re-tokenizes the original name and
   fails to match the split tokens. Contentless mode trusts the inverted
   index directly. Trade-off: highlight()/snippet() unavailable (never used).

Requires full reindex to rebuild FTS5 tables with new schema + tokens.

Author: Your Name

src/pipeline/pipeline.c

@@ -820,12 +820,18 @@ int cbm_pipeline_run(cbm_pipeline_t *p) {
             cbm_log_info("pass.timing", "pass", "persist_hashes", "files", itoa_buf(file_count));
 
             /* Backfill FTS5 index: the direct B-tree dump bypasses SQLite triggers,
-             * so the FTS5 table is empty after indexing. Populate it in bulk now. */
+             * so the FTS5 table is empty after indexing. Populate it in bulk now.
+             * cbm_camel_split(name) splits camelCase into individual tokens so
+             * "updateCloudClient" becomes searchable as "update", "Cloud", "Client". */
             cbm_store_t *fts_store = cbm_store_open_path(db_path);
             if (fts_store) {
+                /* Contentless FTS5 (content='') — use plain INSERT, not INSERT OR REPLACE.
+                 * Clear first to handle reindex scenarios, then bulk insert. */
                 cbm_store_exec(fts_store,
-                    "INSERT OR REPLACE INTO nodes_fts(rowid, name, qualified_name, label, file_path) "
-                    "SELECT id, name, qualified_name, label, file_path FROM nodes;");
+                    "DELETE FROM nodes_fts;");
+                cbm_store_exec(fts_store,
+                    "INSERT INTO nodes_fts(rowid, name, qualified_name, label, file_path) "
+                    "SELECT id, cbm_camel_split(name), qualified_name, label, file_path FROM nodes;");
                 cbm_store_close(fts_store);
             }
 

src/store/store.c

@@ -245,14 +245,19 @@ static int create_user_indexes(cbm_store_t *s) {
     if (rc != SQLITE_OK) return rc;
 
     /* FTS5 full-text search index on node names for BM25 ranking.
-     * content='nodes' makes it an external-content table — synced via triggers.
+     * content='' makes it a contentless table — it stores only the inverted index,
+     * not the original text. This is required for camelCase token splitting:
+     * we index "createSession create Session" but the source table has "createSession".
+     * With content='nodes', FTS5 would re-verify against the source and fail to match
+     * the split tokens. Contentless mode trusts the inverted index directly.
+     * Trade-off: highlight()/snippet() unavailable, but we never use them.
      * Each DDL statement must be executed separately for FTS5 compatibility. */
     {
         char *fts_err = NULL;
         int fts_rc = sqlite3_exec(s->db,
             "CREATE VIRTUAL TABLE IF NOT EXISTS nodes_fts USING fts5("
             "name, qualified_name, label, file_path,"
-            "content='nodes', content_rowid='id',"
+            "content='', content_rowid='id',"
             "tokenize='unicode61 remove_diacritics 2'"
             ");",
             NULL, NULL, &fts_err);
@@ -263,22 +268,30 @@ static int create_user_indexes(cbm_store_t *s) {
         }
     }
 
-    /* Sync triggers: keep FTS index up to date when nodes change */
+    /* Sync triggers: keep FTS index up to date when nodes change.
+     * cbm_camel_split(name) splits camelCase into individual tokens so
+     * "updateCloudClient" is searchable as "update", "Cloud", "Client".
+     * Contentless FTS5 (content='') requires delete operations to provide the
+     * exact same tokenized content that was originally inserted. */
     exec_sql(s, "CREATE TRIGGER IF NOT EXISTS nodes_fts_ai AFTER INSERT ON nodes BEGIN"
                 "  INSERT INTO nodes_fts(rowid, name, qualified_name, label, file_path)"
-                "  VALUES (new.id, new.name, new.qualified_name, new.label, new.file_path);"
+                "  VALUES (new.id, cbm_camel_split(new.name), new.qualified_name,"
+                "          new.label, new.file_path);"
                 "END;");
 
     exec_sql(s, "CREATE TRIGGER IF NOT EXISTS nodes_fts_ad AFTER DELETE ON nodes BEGIN"
                 "  INSERT INTO nodes_fts(nodes_fts, rowid, name, qualified_name, label, file_path)"
-                "  VALUES ('delete', old.id, old.name, old.qualified_name, old.label, old.file_path);"
+                "  VALUES ('delete', old.id, cbm_camel_split(old.name), old.qualified_name,"
+                "          old.label, old.file_path);"
                 "END;");
 
     exec_sql(s, "CREATE TRIGGER IF NOT EXISTS nodes_fts_au AFTER UPDATE ON nodes BEGIN"
                 "  INSERT INTO nodes_fts(nodes_fts, rowid, name, qualified_name, label, file_path)"
-                "  VALUES ('delete', old.id, old.name, old.qualified_name, old.label, old.file_path);"
+                "  VALUES ('delete', old.id, cbm_camel_split(old.name), old.qualified_name,"
+                "          old.label, old.file_path);"
                 "  INSERT INTO nodes_fts(rowid, name, qualified_name, label, file_path)"
-                "  VALUES (new.id, new.name, new.qualified_name, new.label, new.file_path);"
+                "  VALUES (new.id, cbm_camel_split(new.name), new.qualified_name,"
+                "          new.label, new.file_path);"
                 "END;");
 
     return SQLITE_OK;
@@ -338,6 +351,52 @@ static void sqlite_regexp(sqlite3_context *ctx, int argc, sqlite3_value **argv)
     sqlite3_result_int(ctx, rc == 0 ? 1 : 0);
 }
 
+/* CamelCase token splitter for FTS5.
+ * "updateCloudClient" → "updateCloudClient update Cloud Client"
+ * "HTMLParser"        → "HTMLParser HTML Parser"
+ * "getURL"            → "getURL get URL"
+ * Preserves original name as first token for exact-match queries,
+ * then appends space-split words for broad keyword matching. */
+static void sqlite_camel_split(sqlite3_context *ctx, int argc, sqlite3_value **argv) {
+    (void)argc;
+    const char *input = (const char *)sqlite3_value_text(argv[0]);
+    if (!input || !input[0]) {
+        sqlite3_result_text(ctx, input ? input : "", -1, SQLITE_TRANSIENT);
+        return;
+    }
+
+    char buf[2048];
+    /* Start with the original name (preserves exact-match capability) */
+    int len = snprintf(buf, sizeof(buf), "%s ", input);
+
+    /* Walk input, insert space before each camelCase boundary:
+     * - lowercase→Uppercase: "updateCloud" → "update Cloud"
+     * - Uppercase→Uppercase+lowercase: "HTMLParser" → "HTML Parser" */
+    for (int i = 0; input[i] && len < (int)sizeof(buf) - 2; i++) {
+        if (i > 0) {
+            bool split = false;
+            /* lowercase followed by Uppercase: updateC → update C */
+            if (input[i] >= 'A' && input[i] <= 'Z' &&
+                input[i - 1] >= 'a' && input[i - 1] <= 'z') {
+                split = true;
+            }
+            /* Uppercase followed by Uppercase+lowercase: HTMLParser → HTML Parser
+             * Only split before the LAST uppercase in a run */
+            if (input[i] >= 'A' && input[i] <= 'Z' &&
+                input[i - 1] >= 'A' && input[i - 1] <= 'Z' &&
+                input[i + 1] >= 'a' && input[i + 1] <= 'z') {
+                split = true;
+            }
+            if (split) {
+                buf[len++] = ' ';
+            }
+        }
+        buf[len++] = input[i];
+    }
+    buf[len] = '\0';
+    sqlite3_result_text(ctx, buf, len, SQLITE_TRANSIENT);
+}
+
 /* Case-insensitive REGEXP variant */
 static void sqlite_iregexp(sqlite3_context *ctx, int argc, sqlite3_value **argv) {
     (void)argc;
@@ -411,6 +470,9 @@ static cbm_store_t *store_open_internal(const char *path, bool in_memory) {
     /* Case-insensitive variant for search with case_sensitive=false */
     sqlite3_create_function(s->db, "iregexp", 2, SQLITE_UTF8 | SQLITE_DETERMINISTIC, NULL,
                             sqlite_iregexp, NULL, NULL);
+    /* CamelCase splitter for FTS5 indexing — used in triggers and backfill */
+    sqlite3_create_function(s->db, "cbm_camel_split", 1, SQLITE_UTF8 | SQLITE_DETERMINISTIC,
+                            NULL, sqlite_camel_split, NULL, NULL);
 
     if (configure_pragmas(s, in_memory) != CBM_STORE_OK || init_schema(s) != CBM_STORE_OK ||
         create_user_indexes(s) != CBM_STORE_OK) {
@@ -463,6 +525,9 @@ cbm_store_t *cbm_store_open_path_query(const char *db_path) {
                             sqlite_regexp, NULL, NULL);
     sqlite3_create_function(s->db, "iregexp", 2, SQLITE_UTF8 | SQLITE_DETERMINISTIC, NULL,
                             sqlite_iregexp, NULL, NULL);
+    /* CamelCase splitter for FTS5 — must be registered before triggers fire */
+    sqlite3_create_function(s->db, "cbm_camel_split", 1, SQLITE_UTF8 | SQLITE_DETERMINISTIC,
+                            NULL, sqlite_camel_split, NULL, NULL);
 
     if (configure_pragmas(s, false) != CBM_STORE_OK) {
         sqlite3_close(s->db);
@@ -2060,9 +2125,12 @@ int cbm_store_search(cbm_store_t *s, const cbm_search_params_t *params, cbm_sear
         }
 
         char fts_sql[4096];
-        /* Join with FTS5 table, filter by project/label, order by BM25 rank.
-         * Exclude noise labels (File, Folder, Module, Section, Variable, Project)
-         * and boost Function/Method/Class via a structural score added to BM25. */
+        /* Join with FTS5 table, filter by project/label, order by pure BM25 relevance.
+         * Exclude noise labels (File, Folder, Module, Section, Variable, Project).
+         * Label-type boost: prefer Functions/Methods/Routes/Classes over generic nodes.
+         * No fan_in/popularity boost — that corrupts relevance for discovery queries
+         * (e.g. "update" with fan_in=222 would outrank the actually relevant match).
+         * in_deg/out_deg are still returned for display but do NOT affect ranking. */
         int flen = snprintf(fts_sql, sizeof(fts_sql),
             "SELECT n.id, n.project, n.label, n.name, n.qualified_name, "
             "n.file_path, n.start_line, n.end_line, n.properties, "
@@ -2073,7 +2141,6 @@ int cbm_store_search(cbm_store_t *s, const cbm_search_params_t *params, cbm_sear
             "        WHEN n.label IN ('Class','Interface','Type') THEN 5.0 "
             "        WHEN n.label = 'Route' THEN 8.0 "
             "        ELSE 0.0 END "
-            " - CASE WHEN (SELECT COUNT(*) FROM edges e WHERE e.target_id = n.id AND e.type = 'CALLS') > 5 THEN 3.0 ELSE 0.0 END"
             ") AS rank "
             "FROM nodes_fts "
             "JOIN nodes n ON n.id = nodes_fts.rowid "