fix: symmetric umlaut/ß folding in lexical path + eval digraph case

Smoke-testing invented German dictations exposed that `ae`/`oe`/`ue`
spellings collapsed lexical recall (e.g. "Hyperlipidaemie" → garbage,
"Hyperlipidämie" → correct). normalize() now folds ä→ae, ö→oe, ü→ue,
ß→ss; it is applied only in lexical.rs (both corpus index and query),
so the path is symmetric under either spelling. The embedding path
(raw text for bge-m3) is intentionally untouched.

eval.rs inject_errors case 2 changed from umlaut-deletion (ä→a) to the
realistic ASCII-digraph variant (ä→ae) so the eval harness exercises —
and guards against regression of — exactly this failure class.

TDD: failing tests first; normalize_tests 5/5, full suite 44/0.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

src/normalize.rs

@@ -9,8 +9,28 @@ const ABBREV: &[(&str, &str)] = &[
     ("li.", "links"),
 ];
 
+/// Fold German umlauts and eszett to their ASCII digraph equivalents so that
+/// "Hyperlipidämie" and "Hyperlipidaemie" collapse to one canonical token.
+/// Operates on already-lowercased text (so only the lowercase forms appear).
+/// Applied to BOTH the indexed corpus text and the query (see lexical.rs), so
+/// the lexical path stays symmetric under either spelling. NOT applied to the
+/// embedding path — bge-m3 is multilingual and handles umlauts natively.
+fn fold_umlauts(s: &str) -> String {
+    let mut out = String::with_capacity(s.len());
+    for c in s.chars() {
+        match c {
+            'ä' => out.push_str("ae"),
+            'ö' => out.push_str("oe"),
+            'ü' => out.push_str("ue"),
+            'ß' => out.push_str("ss"),
+            _ => out.push(c),
+        }
+    }
+    out
+}
+
 pub fn normalize(input: &str) -> String {
-    let lower = input.to_lowercase();
+    let lower = fold_umlauts(&input.to_lowercase());
     let mut out: Vec<String> = Vec::new();
     for tok in lower.split_whitespace() {
         let replaced = ABBREV.iter()