fix: harden against confirmed review findings

pixi.toml

@@ -11,4 +11,4 @@ test = "mojo run -I src test/test_unicodedata.mojo"
 conformance = "mojo run -I src test/conformance.mojo"
 
 [dependencies]
-mojo = ">=1.0.0b3,<2"
+mojo = ">=1.0.0b3.dev0,<2"

test/conformance.mojo

@@ -11,12 +11,19 @@ forms holds (UAX #15, NormalizationTest.txt header):
     NFKC: c4 == toNFKC(c1) == toNFKC(c2) == toNFKC(c3) == toNFKC(c4) == toNFKC(c5)
     NFKD: c5 == toNFKD(c1) == toNFKD(c2) == toNFKD(c3) == toNFKD(c4) == toNFKD(c5)
 
-This is a scoreboard, not a gate: it always exits successfully. Pass the
-name of a part (e.g. "@Part1") as an argument to print the first few
-failing rows for that part.
+This is a CI gate: it exits non-zero if any row fails, or if the corpus
+does not contain exactly EXPECTED_ROWS rows (a guard against a truncated or
+swapped data file silently shrinking coverage). Pass the name of a part
+(e.g. "@Part1") as an argument to print the first few failing rows for that
+part.
 """
 
-from std.sys import argv
+from std.sys import argv, exit
+
+# The official UAX #15 NormalizationTest.txt corpus vendored in test/data has
+# exactly this many data rows. If the file is truncated, replaced, or fails to
+# parse, the row count drifts and coverage silently shrinks -- so we assert it.
+comptime EXPECTED_ROWS = 20034
 
 from unicodedata import Normalizer
 
@@ -191,6 +198,21 @@ def main() raises:
     print("-----------------------------------------")
     print("TOTAL:", total_pass, "/", total, "rows passing")
 
+    # Real CI gate. Exit non-zero on any failure so a regression cannot ship
+    # green, and verify the corpus is intact so coverage cannot silently shrink.
+    var failed = False
+    if total != EXPECTED_ROWS:
+        print(
+            "GATE FAIL: parsed", total, "rows but expected", EXPECTED_ROWS,
+            "-- corpus is truncated, swapped, or misparsed.",
+        )
+        failed = True
+    if total_pass != total:
+        print("GATE FAIL:", total - total_pass, "row(s) failed conformance.")
+        failed = True
+    if failed:
+        exit(1)
+
 
 def _dump(s: String) -> String:
     """Hex codepoint dump of a string, for failure diagnostics."""

test/test_unicodedata.mojo

@@ -149,5 +149,29 @@ def test_normalizer_reuse() raises:
     assert_equal(norm.casefold(String("ẞ")), String("ss"))
 
 
+def test_conformance_corpus_intact() raises:
+    """Regression guard for the conformance gate.
+
+    The conformance harness (test/conformance.mojo) fails CI unless the corpus
+    parses to exactly 20034 rows. This mirrors that row count here so a
+    truncated, swapped, or misparsed NormalizationTest.txt -- which would
+    silently shrink coverage while still showing 100% -- fails the unit suite
+    too, independently of the harness.
+    """
+    var raw = open("test/data/normalization_test.txt", "r").read()
+    var lines = raw.split("\n")
+    var rows = 0
+    for li in range(len(lines)):
+        var line = String(StringSlice(lines[li]).strip())
+        if line.byte_length() == 0:
+            continue
+        if line.startswith("@"):
+            continue
+        if len(line.split(";")) < 5:
+            continue
+        rows += 1
+    assert_equal(rows, 20034)
+
+
 def main() raises:
     TestSuite.discover_tests[__functions_in_module()]().run()