Author: Andy Green <andy@warmcat.com>
Date: Sat Oct 03 10:24:57 2026 +0100

npro-core, npro-test: let Miri run the tests, without the heaviest

Miri interprets the tests for a target sai has no builder for, a
big-endian or a 32-bit one, which is where byte order and the size of
usize could make the same code behave differently.  It is about a
thousand times slower than native.  Four tests do bulk work that byte
order and word size do not bear on:
- the exhaustive UTF-8 sweep;
- SHA-1 of a million bytes;
- parsing 4097 steps of transcript;
- reading every vendored transcript.

Those four are ignored under cfg(miri), with the reason given.  Native
runs keep all of them, and the million-byte SHA-1 vector moves to a test
of its own so that the other RFC 3174 vectors still run under Miri.

With that, the whole suite runs under Miri in about two minutes per
target: 31 tests, all passing, for s390x (big-endian) and i686 (32-bit).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019kg5Eemy68ZaqDBcUJQG6J


diff --git a/crates/npro-core/src/sha1.rs b/crates/npro-core/src/sha1.rs
index 76913f0..7c0fa33 100644
--- a/crates/npro-core/src/sha1.rs
+++ b/crates/npro-core/src/sha1.rs
@@ -189,16 +189,24 @@ mod tests {
             "84983e441c3bd26ebaae4aa1f95129e5e54670f1"
         );
         assert_eq!(
-            hex(Sha1::digest(&vec![b'a'; 1_000_000])),
-            "34aa973cd4c4daa4f61eeb2bdbad27316534016f"
-        );
-        assert_eq!(
             hex(Sha1::digest(b"")),
             "da39a3ee5e6b4b0d3255bfef95601890afd80709"
         );
     }
 
     #[test]
+    #[cfg_attr(
+        miri,
+        ignore = "a million bytes: native runs keep it, Miri runs the shorter vectors"
+    )]
+    fn rfc_3174_a_million_a() {
+        assert_eq!(
+            hex(Sha1::digest(&vec![b'a'; 1_000_000])),
+            "34aa973cd4c4daa4f61eeb2bdbad27316534016f"
+        );
+    }
+
+    #[test]
     fn padding_at_every_block_boundary() {
         // lengths around 55 / 56 / 64 bytes put the 0x80 and the length in
         // the last block or the next one: digests of 'a' repeated n times,
diff --git a/crates/npro-core/src/utf8.rs b/crates/npro-core/src/utf8.rs
index 0a7f14e..155e541 100644
--- a/crates/npro-core/src/utf8.rs
+++ b/crates/npro-core/src/utf8.rs
@@ -149,6 +149,10 @@ mod tests {
     }
 
     #[test]
+    #[cfg_attr(
+        miri,
+        ignore = "millions of cases: native runs keep it, Miri runs the targeted ones"
+    )]
     fn agrees_with_the_standard_library_on_every_short_sequence() {
         for a in 0..=255u8 {
             assert_eq!(whole(&[a]), oracle(&[a]), "{a:02x}");
diff --git a/crates/npro-test/src/transcript.rs b/crates/npro-test/src/transcript.rs
index a23d527..4171ad1 100644
--- a/crates/npro-test/src/transcript.rs
+++ b/crates/npro-test/src/transcript.rs
@@ -621,6 +621,10 @@ mod tests {
     }
 
     #[test]
+    #[cfg_attr(
+        miri,
+        ignore = "megabytes of text: native runs keep it, Miri runs the small cases"
+    )]
     fn refuses_too_many_steps_and_too_much_text() {
         let many = vec![r#"{"t": 0, "rx": ""}"#; MAX_STEPS + 1].join(",");
         assert!(matches!(refused(&with_steps(&many)), Error::TooManySteps));
diff --git a/crates/npro-test/tests/vendored.rs b/crates/npro-test/tests/vendored.rs
index 2cd4b2d..d95ed0b 100644
--- a/crates/npro-test/tests/vendored.rs
+++ b/crates/npro-test/tests/vendored.rs
@@ -3,6 +3,10 @@
 use npro_test::{Side, StepKind, vendored};
 
 #[test]
+#[cfg_attr(
+    miri,
+    ignore = "hundreds of kilobytes of transcripts: native runs keep it"
+)]
 fn every_vendored_transcript_reads() {
     let all = vendored().unwrap();