labels: reader-backed detection, Criterion fix, menu-language fallback

- DetectFn now takes a SectorSource so a parser can inspect a jar's
  central directory in detect() instead of firing on "any BD-J jar".
  dbp/deluxe do the real com/<vendor>/ prefix check up front, so each
  claims only its own discs (foundational for scaling the registry).
- criterion: treat a stream-map value of 0 as unmapped and synthesize a
  real 1-based number, so a 0 can't shadow or collide with a genuine
  stream 1 (with regression tests).
- png_filenames: new Low-confidence, last-resort parser reading menu
  language from {title}_UHD01_{LANG}_Composite artwork; sits below the
  MPLS floor so a real framework parser always wins.
- vocab: add menu_lang() for 639-2/B to 639-2/T menu-token normalization.

src/labels/dbp.rs

@@ -35,14 +35,16 @@ use crate::sector::SectorSource;
 use crate::udf::UdfFs;
 use std::collections::BTreeMap;
 
-/// dbp detect can't peek inside a jar without a SectorSource (the
-/// trait function only takes `&UdfFs`), so we trigger on the cheap
-/// signal "any top-level .jar in /BDMV/JAR/." That fires on every
-/// BD-J disc, but parse() does the real `com/dbp/` check and
-/// returns None on a mismatch — so this parser only ever consumes
-/// time on discs that fell through every earlier parser.
-pub fn detect(udf: &UdfFs) -> bool {
-    jar::has_any_top_level_jar(udf)
+/// The real dbp signal is the `com/dbp/` package prefix inside a top-level
+/// jar's central directory. With a reader in `detect`, we check that directly
+/// (a cheap central-directory scan, no class decode) so this parser claims
+/// only dbp discs instead of firing on every BD-J disc. `parse()` repeats the
+/// check as belt-and-suspenders.
+pub fn detect(reader: &mut dyn SectorSource, udf: &UdfFs) -> bool {
+    jar::for_each_jar(reader, udf, |_entry, archive| {
+        jar::has_path_prefix(archive, "com/dbp/").then_some(())
+    })
+    .is_some()
 }
 
 /// Scan every top-level `/BDMV/JAR/*.jar` for the dbp framework and