Merge pull request #2685 from storytold/land/1937

Round-trip harness and corpus report: open, save, reopen and compare (#1937, E12)

apps/cadcraft-cli/src/args.rs

@@ -81,6 +81,7 @@ pub(crate) enum Command {
     Mcp(McpArgs),
     Perf(PerfArgs),
     Sample(SampleArgs),
+    Roundtrip(RoundtripArgs),
 }
 
 /// Print a drawing's summary as JSON
@@ -308,6 +309,40 @@ fn perf_target(p: PathBuf) -> PerfTarget {
     }
 }
 
+/// Open each DXF/DWG, save it as DXF and DWG, reopen and compare; writes report.md and report.json
+///
+/// FOLDER defaults to $CADCRAFT_CORPUS_DIR. With CADCRAFT_CORPUS_STRICT=1 it fails when a file
+/// doesn't open, save or reopen, or loses records on save.
+#[derive(Args)]
+pub(crate) struct RoundtripArgs {
+    /// A drawing, or a folder of them
+    #[arg(value_name = "FILE|FOLDER")]
+    pub(crate) path: Option<PathBuf>,
+    /// Where the report goes (default: `report` beside the files)
+    #[arg(long, value_name = "DIR")]
+    pub(crate) out: Option<PathBuf>,
+    /// Keep the saved files, in `saved` under the report folder
+    #[arg(long)]
+    pub(crate) keep_saved: bool,
+    /// Save only in each file's own format
+    #[arg(long)]
+    pub(crate) same_format: bool,
+    /// Don't count what each file holds (faster; no "read lost" column)
+    #[arg(long)]
+    pub(crate) no_census: bool,
+    /// How far numbers may move and still count as equal
+    #[arg(long, value_name = "T", default_value_t = cadcraft_io::roundtrip::Options::default().abs_tol, value_parser = tolerance)]
+    pub(crate) tolerance: f64,
+}
+
+/// `--tolerance`: a finite number, 0 or more.
+fn tolerance(v: &str) -> Result<f64, ValueError> {
+    match v.parse::<f64>() {
+        Ok(t) if t.is_finite() && t >= 0.0 => Ok(t),
+        _ => Err(ValueError::Tolerance(v.to_string())),
+    }
+}
+
 /// Write a built-in sample drawing
 #[derive(Args)]
 pub(crate) struct SampleArgs {
@@ -360,6 +395,8 @@ pub(crate) enum ValueError {
     CmdId,
     #[error(transparent)]
     CmdJson(#[from] serde_json::Error),
+    #[error("expected a number of 0 or more, got `{0}`")]
+    Tolerance(String),
 }
 
 /// Why `parse` refuses flags clap accepted: clap prints it as a usage error (exit 2).

apps/cadcraft-cli/src/error.rs

@@ -1,5 +1,6 @@
 use crate::output::OutputError;
 use crate::perf::PerfError;
+use crate::roundtrip::RoundtripError;
 use crate::run::RunError;
 
 /// Why a subcommand failed: printed as `error: {e}`, exit 1.
@@ -19,6 +20,8 @@ pub(crate) enum CliError {
     Output(#[from] OutputError),
     #[error(transparent)]
     Perf(#[from] PerfError),
+    #[error(transparent)]
+    Roundtrip(#[from] RoundtripError),
     /// The engine and io crates take file names as text.
     #[error("{}: the file name is not valid Unicode", .0.display())]
     NotUnicode(std::path::PathBuf),

apps/cadcraft-cli/src/lib.rs

@@ -16,6 +16,7 @@ mod info;
 mod mcp;
 mod output;
 mod perf;
+mod roundtrip;
 mod run;
 mod sample;
 
@@ -32,6 +33,7 @@ use output::Streams;
 #[cfg(test)]
 use output::{export_window, model_views, write_atomic};
 use perf::perf;
+use roundtrip::roundtrip;
 use run::run_cmds;
 #[cfg(test)]
 use run::{fill_last, returned_handle};
@@ -99,7 +101,7 @@ where
         Command::Run(a) => a.settings.apply(),
         Command::Mcp(a) => a.settings.apply(),
         Command::Sample(a) => a.settings.apply(),
-        Command::Info(_) | Command::Commands(_) | Command::Perf(_) => {}
+        Command::Info(_) | Command::Commands(_) | Command::Perf(_) | Command::Roundtrip(_) => {}
     }
     let mut io = Streams::new(out, err);
     let result = match &cli.command {
@@ -110,6 +112,7 @@ where
         Command::Mcp(a) => mcp(a),
         Command::Perf(a) => perf(a.target.as_ref(), &mut io),
         Command::Sample(a) => sample(a, &mut io),
+        Command::Roundtrip(a) => roundtrip(a, &mut io),
     };
     let mut code = 0;
     if let Err(e) = result {

apps/cadcraft-cli/src/roundtrip.rs

@@ -0,0 +1,93 @@
+//! `roundtrip`: the corpus harness of `cadcraft_io::roundtrip::corpus` over a file or folder.
+
+use super::args::RoundtripArgs;
+use super::error::{CliError, Result};
+use super::output::Streams;
+
+/// Why `roundtrip` stopped.
+#[derive(Debug, thiserror::Error)]
+pub(crate) enum RoundtripError {
+    #[error("roundtrip: give a FILE or FOLDER, or set CADCRAFT_CORPUS_DIR")]
+    NoPath,
+    #[error("roundtrip: no such file or folder: {}", .0.display())]
+    Missing(std::path::PathBuf),
+    #[error("roundtrip: no DXF or DWG files in {}", .0.display())]
+    Empty(std::path::PathBuf),
+    #[error("roundtrip: can't write the report to {}: {source}", .path.display())]
+    Report { path: std::path::PathBuf, source: std::io::Error },
+    #[error("roundtrip: {} strict failures:\n{}", .0.len(), .0.join("\n"))]
+    Strict(Vec<String>),
+}
+
+/// What the walk couldn't look at, so the report isn't taken for the whole corpus: said on
+/// stderr, and failures of a strict run.
+fn left_out(w: &cadcraft_io::roundtrip::corpus::Walk) -> Vec<String> {
+    let mut out: Vec<String> = w.unreadable.iter().map(|u| format!("not read: {u}")).collect();
+    if w.truncated {
+        out.push("left out: files or folders past the harness limits".to_string());
+    }
+    out
+}
+
+pub(crate) fn roundtrip(a: &RoundtripArgs, io: &mut Streams) -> Result {
+    use cadcraft_io::roundtrip::corpus;
+    let path = a.path.clone().or_else(|| std::env::var_os("CADCRAFT_CORPUS_DIR").map(Into::into)).ok_or(RoundtripError::NoPath)?;
+    let mut missed = Vec::new();
+    let (root, files, out) = if path.is_dir() {
+        let out = a.out.clone().unwrap_or_else(|| path.join("report"));
+        // The report folder, and the saved copies in it, are never read back as corpus files.
+        let walk = corpus::walk(&path, &[out.clone(), out.join("saved")]);
+        missed = left_out(&walk);
+        for m in &missed {
+            io.errln(format_args!("roundtrip: {m}"));
+        }
+        (path, walk.files, out)
+    } else if path.is_file() {
+        let root = path.parent().map(std::path::Path::to_path_buf).unwrap_or_default();
+        let out = a.out.clone().unwrap_or_else(|| root.join("report"));
+        (root, vec![path], out)
+    } else {
+        return Err(RoundtripError::Missing(path).into());
+    };
+    if files.is_empty() {
+        // A folder it couldn't read isn't an empty one.
+        return Err(if missed.is_empty() { RoundtripError::Empty(root) } else { RoundtripError::Strict(missed) }.into());
+    }
+    let mut opts = corpus::CorpusOptions { cross_format: !a.same_format, census: !a.no_census, ..Default::default() };
+    opts.compare.abs_tol = a.tolerance;
+    if a.keep_saved {
+        opts.keep_saved = Some(out.join("saved"));
+    }
+    let report = corpus::run(&root, &files, &opts, &mut |i, n, rel| eprintln!("[{}/{n}] {rel}", i + 1));
+    for f in &report.files {
+        let routes: Vec<String> = f
+            .routes
+            .iter()
+            .map(|r| match (&r.save_error, &r.reopen_error) {
+                (Some(_), _) => format!("{}: save failed", r.route),
+                (None, Some(_)) => format!("{}: reopen failed", r.route),
+                (None, None) => {
+                    let t = r.comparison.total();
+                    format!("{}: {} lost, {} added, {} changed", r.route, t.lost, t.added, t.changed)
+                }
+            })
+            .collect();
+        match &f.open_error {
+            Some(e) => io.outln(format_args!("{}: open failed: {e}", f.path)),
+            None => io.outln(format_args!("{}: opened; {}", f.path, routes.join("; "))),
+        }
+    }
+    let (md, json) = report.write(&out).map_err(|source| RoundtripError::Report { path: out.clone(), source })?;
+    io.outln(format_args!("report: {}", md.display()));
+    io.outln(format_args!("        {}", json.display()));
+    // CADCRAFT_CORPUS_STRICT=1: a file that doesn't open, save or reopen, or records a writer
+    // leaves out, fail the run (the report is written first).
+    if std::env::var("CADCRAFT_CORPUS_STRICT").is_ok_and(|v| v == "1") {
+        let mut failed = report.strict_failures();
+        failed.extend(missed);
+        if !failed.is_empty() {
+            return Err(CliError::from(RoundtripError::Strict(failed)));
+        }
+    }
+    Ok(())
+}

apps/cadcraft-cli/tests/man_and_completions.rs

@@ -20,7 +20,7 @@ fn the_example_writes_the_files_the_packages_install() {
     let mut written: Vec<String> = std::fs::read_dir(&dir).unwrap().map(|e| e.unwrap().file_name().into_string().unwrap()).collect();
     written.sort();
     let completions = ["cadcraft-cli.bash", "_cadcraft-cli", "cadcraft-cli.fish", "cadcraft-cli.elv", "_cadcraft-cli.ps1"];
-    let subcommands = ["info", "convert", "run", "commands", "mcp", "perf", "sample"];
+    let subcommands = ["info", "convert", "run", "commands", "mcp", "perf", "sample", "roundtrip"];
     let mut expected: Vec<String> = completions.map(String::from).into();
     expected.push("cadcraft-cli.1".into());
     expected.extend(subcommands.iter().map(|s| format!("cadcraft-cli-{s}.1")));

crates/dwg/src/lib.rs

@@ -65,7 +65,7 @@ fn mb(n: u64) -> String {
 #[cfg(not(no_dwg_bridge))]
 pub mod worker;
 #[cfg(not(no_dwg_bridge))]
-pub use worker::{WORKER_ARG, crash_worker_for_test, decode, encode_as, serve_if_asked, use_own_worker, use_worker};
+pub use worker::{WORKER_ARG, crash_worker_for_test, decode, encode_as, encode_noted, serve_if_asked, use_own_worker, use_worker};
 
 #[cfg(not(no_dwg_bridge))]
 mod native {
@@ -544,15 +544,22 @@ mod native {
     /// Leave out what acadrust can't write as this DWG version: other programs' custom entities
     /// and objects without raw DWG data of that version. acadrust 0.6.4 refuses the whole file
     /// for one of them (0.6.3 left them out); CADCraft already tells the user which entity types
-    /// a DWG doesn't keep (`not_saved`), so the rest of the drawing is written. Returns how many.
-    fn drop_unwritable(doc: &mut acadrust::CadDocument) -> usize {
+    /// a DWG doesn't keep (`not_saved`), so the rest of the drawing is written. Returns how many
+    /// records of each DXF type were left out, so a save can say so.
+    fn drop_unwritable(doc: &mut acadrust::CadDocument) -> super::Unwritten {
         let version = doc.version;
+        let mut dropped = super::Unwritten::new();
+        let mut count = |name: &str| {
+            let n = dropped.entry(name.trim().to_ascii_uppercase()).or_insert(0);
+            *n = n.saturating_add(1);
+        };
         let unwritable: Vec<acadrust::Handle> = doc
             .entities()
             .filter_map(|e| match e {
                 acadrust::entities::EntityType::Unknown(u)
                     if u.raw_dwg_data.is_none() || u.dwg_source_version.is_some_and(|v| v != version) || version == acadrust::DxfVersion::AC1009 =>
                 {
+                    count(&u.dxf_name);
                     Some(u.common.handle)
                 }
                 _ => None,
@@ -565,9 +572,10 @@ mod native {
             .objects
             .iter()
             .filter_map(|(h, o)| match o {
-                acadrust::objects::ObjectType::Unknown { raw_dwg_data, raw_dwg_version, .. }
+                acadrust::objects::ObjectType::Unknown { type_name, raw_dwg_data, raw_dwg_version, .. }
                     if raw_dwg_data.is_none() || raw_dwg_version.is_some_and(|v| v != version) || version == acadrust::DxfVersion::AC1009 =>
                 {
+                    count(type_name);
                     Some(*h)
                 }
                 _ => None,
@@ -576,7 +584,7 @@ mod native {
         for h in &objects {
             doc.objects.remove(h);
         }
-        unwritable.len() + objects.len()
+        dropped
     }
 
     /// Header variables the DWG reader fills in but acadrust's DXF writer (0.6.3) leaves out of
@@ -644,12 +652,19 @@ mod native {
     /// Convert DXF bytes into a DWG file of the DXF's version, and at least AutoCAD 2004
     /// (AC1018).
     pub fn dxf_to_dwg(dxf: &[u8]) -> Result<Vec<u8>, String> {
-        convert(dxf, None)
+        convert(dxf, None).map(|(dwg, _)| dwg)
     }
 
     /// Convert DXF bytes into a DWG file of `version`, a code from [`super::VERSIONS`] ("AC1032"),
     /// whatever version the DXF declares.
     pub fn dxf_to_dwg_as(dxf: &[u8], version: &str) -> Result<Vec<u8>, String> {
+        dxf_to_dwg_noted(dxf, version).map(|(dwg, _)| dwg)
+    }
+
+    /// [`dxf_to_dwg_as`], also returning what the DWG writer left out: how many records of each
+    /// DXF type it can't write (other programs' custom entities and objects that the DWG reader
+    /// didn't read from a DWG of this version), empty when the DWG keeps everything.
+    pub fn dxf_to_dwg_noted(dxf: &[u8], version: &str) -> Result<(Vec<u8>, super::Unwritten), String> {
         let v = match version.trim().to_ascii_uppercase().as_str() {
             "AC1015" => acadrust::DxfVersion::AC1015,
             "AC1018" => acadrust::DxfVersion::AC1018,
@@ -707,7 +722,7 @@ mod native {
         }
     }
 
-    fn convert(dxf: &[u8], version: Option<acadrust::DxfVersion>) -> Result<Vec<u8>, String> {
+    fn convert(dxf: &[u8], version: Option<acadrust::DxfVersion>) -> Result<(Vec<u8>, super::Unwritten), String> {
         // Reject a pathologically large DXF before handing it to the third-party decoder, which
         // allocates tables and section buffers from the input's declared sizes and cannot be
         // interrupted once it starts (CWE-789). This bounds the large-input variant; the decoder
@@ -732,21 +747,23 @@ mod native {
             None => {}
         }
         // The fix-ups run on the third-party reader's document, so a panic in one is an error, not a crash.
-        guarded(AssertUnwindSafe(|| {
+        let unwritten = guarded(AssertUnwindSafe(|| {
             super::fail_in_test("convert");
             other_layout_owners(&mut doc);
             table_style_rows(&mut doc, true);
             dimension_flags(&mut doc);
-            drop_unwritable(&mut doc);
+            let unwritten = drop_unwritable(&mut doc);
             current_dimstyle(&mut doc);
             add_variable_dictionary(&mut doc);
             current_text_style(&mut doc);
             current_layer(&mut doc);
+            unwritten
         }))
         .map_err(|_| "the DXF to DWG conversion failed on this file".to_string())?;
-        std::panic::catch_unwind(AssertUnwindSafe(|| acadrust::DwgWriter::write_to_vec(&doc)))
+        let dwg = std::panic::catch_unwind(AssertUnwindSafe(|| acadrust::DwgWriter::write_to_vec(&doc)))
             .map_err(|_| "the DWG writer failed on this drawing".to_string())?
-            .map_err(|e| format!("DWG write: {e}"))
+            .map_err(|e| format!("DWG write: {e}"))?;
+        Ok((dwg, unwritten))
     }
 
     /// AutoCAD's entry name for the drawing variables dictionary (the DXF Reference calls it the
@@ -903,7 +920,11 @@ mod native {
 }
 
 #[cfg(not(no_dwg_bridge))]
-pub use native::{dwg_to_dxf, dwg_to_dxf_noted, dwg_to_dxf_with, dxf_to_dwg, dxf_to_dwg_as};
+pub use native::{dwg_to_dxf, dwg_to_dxf_noted, dwg_to_dxf_with, dxf_to_dwg, dxf_to_dwg_as, dxf_to_dwg_noted};
+
+/// What saving as DWG left out: how many records of each DXF type (upper case) the DWG writer
+/// can't write.
+pub type Unwritten = std::collections::BTreeMap<String, usize>;
 
 // Test-only: the stage ("post_process" or "convert") whose guard the current test checks. The
 // stage then panics on entry, so a test through the public API shows whether it is guarded.
@@ -950,6 +971,10 @@ pub fn decode(bytes: &[u8], limits: Limits) -> Result<(Vec<u8>, Vec<String>), St
 pub fn encode_as(dxf: &[u8], version: &str) -> Result<Vec<u8>, String> {
     dxf_to_dwg_as(dxf, version)
 }
+#[cfg(no_dwg_bridge)]
+pub fn encode_noted(dxf: &[u8], version: &str) -> Result<(Vec<u8>, Unwritten), String> {
+    dxf_to_dwg_noted(dxf, version)
+}
 
 #[cfg(no_dwg_bridge)]
 pub fn dwg_to_dxf(_bytes: &[u8]) -> Result<Vec<u8>, String> {
@@ -971,6 +996,10 @@ pub fn dxf_to_dwg(_dxf: &[u8]) -> Result<Vec<u8>, String> {
 pub fn dxf_to_dwg_as(_dxf: &[u8], _version: &str) -> Result<Vec<u8>, String> {
     Err(format!("DWG files can't be written in the {NO_BRIDGE_BUILD} build yet; save as DXF"))
 }
+#[cfg(no_dwg_bridge)]
+pub fn dxf_to_dwg_noted(_dxf: &[u8], _version: &str) -> Result<(Vec<u8>, Unwritten), String> {
+    Err(format!("DWG files can't be written in the {NO_BRIDGE_BUILD} build yet; save as DXF"))
+}
 
 #[cfg(test)]
 mod tests {
@@ -1302,4 +1331,47 @@ mod tests {
             assert_eq!(read.dwg_tail_byte, 0, "{version}");
         }
     }
+
+    /// Saving says what the DWG writer left out, per type: entities of types acadrust's DXF
+    /// reader doesn't model (another program's custom entity) can't be written; a drawing the DWG
+    /// keeps whole reports nothing.
+    #[cfg(not(no_dwg_bridge))]
+    #[test]
+    fn saving_lists_the_types_the_dwg_writer_leaves_out() {
+        let line = "  0\nLINE\n  8\n0\n100\nAcDbEntity\n100\nAcDbLine\n 10\n0.0\n 20\n0.0\n 30\n0.0\n 11\n10.0\n 21\n0.0\n 31\n0.0\n";
+        let widget = "  0\nACME_WIDGET\n  8\n0\n100\nAcDbEntity\n100\nAcmeWidget\n 10\n1.0\n 20\n2.0\n 30\n0.0\n";
+        let dxf = |body: &str| {
+            format!("  0\nSECTION\n  2\nHEADER\n  9\n$ACADVER\n  1\nAC1032\n  0\nENDSEC\n  0\nSECTION\n  2\nENTITIES\n{body}  0\nENDSEC\n  0\nEOF\n")
+        };
+        for version in ["AC1015", "AC1018", "AC1032"] {
+            let (dwg, unwritten) = dxf_to_dwg_noted(dxf(&format!("{line}{widget}{widget}")).as_bytes(), version).unwrap();
+            assert!(is_dwg(&dwg), "{version}");
+            assert_eq!(unwritten, [("ACME_WIDGET".to_string(), 2)].into_iter().collect::<Unwritten>(), "{version}");
+        }
+        let (_, unwritten) = dxf_to_dwg_noted(dxf(line).as_bytes(), "AC1032").unwrap();
+        assert!(unwritten.is_empty(), "{unwritten:?}");
+    }
+
+    /// An OLE2FRAME, which CADCraft keeps as its DXF groups without modelling it, is written to
+    /// the DWG with its data: it isn't in the list of what the writer left out.
+    #[cfg(not(no_dwg_bridge))]
+    #[test]
+    fn ole_frames_are_written_with_their_data() {
+        let ole = "  0\nOLE2FRAME\n  8\n0\n100\nAcDbEntity\n100\nAcDbOle2Frame\n 70\n2\n 10\n1.0\n 20\n9.0\n 30\n0.0\n 11\n8.0\n 21\n2.0\n 31\n0.0\n 71\n2\n 72\n0\n 73\n1\n 90\n4\n310\nDEADBEEF\n  1\nOLE\n";
+        let dxf =
+            format!("  0\nSECTION\n  2\nHEADER\n  9\n$ACADVER\n  1\nAC1032\n  0\nENDSEC\n  0\nSECTION\n  2\nENTITIES\n{ole}  0\nENDSEC\n  0\nEOF\n");
+        for version in ["AC1015", "AC1018", "AC1032"] {
+            let (dwg, unwritten) = dxf_to_dwg_noted(dxf.as_bytes(), version).unwrap();
+            assert!(unwritten.is_empty(), "{version} {unwritten:?}");
+            let back = acadrust::DwgReader::from_stream(std::io::Cursor::new(dwg)).read().unwrap();
+            let frames: Vec<Vec<u8>> = back
+                .entities()
+                .filter_map(|e| match e {
+                    acadrust::entities::EntityType::Ole2Frame(o) => Some(o.encoded_payload()),
+                    _ => None,
+                })
+                .collect();
+            assert_eq!(frames, [vec![0xDE, 0xAD, 0xBE, 0xEF]], "{version}");
+        }
+    }
 }

crates/dwg/src/worker.rs

@@ -82,15 +82,31 @@ pub fn decode(bytes: &[u8], limits: Limits) -> Result<(Vec<u8>, Vec<String>), St
 
 /// [`super::dxf_to_dwg_as`] in the worker, when there is one.
 pub fn encode_as(dxf: &[u8], version: &str) -> Result<Vec<u8>, String> {
+    encode_noted(dxf, version).map(|(dwg, _)| dwg)
+}
+
+/// [`super::dxf_to_dwg_noted`] in the worker, when there is one. What the writer left out comes
+/// back as notes of `TYPE<tab>count`.
+pub fn encode_noted(dxf: &[u8], version: &str) -> Result<(Vec<u8>, super::Unwritten), String> {
     if u8::try_from(version.len()).is_err() {
         return Err(super::unsupported(version));
     }
     match call(&request(ENCODE, Limits::DEFAULT, version), dxf, MAX_DWG_REPLY, "writing this file") {
-        Some(reply) => reply?.map(|(dwg, _)| dwg),
-        None => super::dxf_to_dwg_as(dxf, version),
+        Some(reply) => reply?.map(|(dwg, notes)| (dwg, unwritten_from(&notes))),
+        None => super::dxf_to_dwg_noted(dxf, version),
     }
 }
 
+fn unwritten_from(notes: &[String]) -> super::Unwritten {
+    notes.iter().filter_map(|n| n.split_once('\t')).filter_map(|(kind, n)| Some((kind.to_string(), n.parse::<usize>().ok()?))).fold(
+        super::Unwritten::new(),
+        |mut m, (kind, n)| {
+            *m.entry(kind).or_insert(0) = n;
+            m
+        },
+    )
+}
+
 /// Make the worker abort, as the third-party code can: what the tests prove isolation with.
 #[doc(hidden)]
 pub fn crash_worker_for_test() -> Result<(), String> {
@@ -241,7 +257,7 @@ fn handle(mut b: &[u8]) -> Reply {
     let version = take(&mut b, usize::from(version_len)).and_then(|v| std::str::from_utf8(v).ok()).ok_or_else(bad)?;
     match op {
         DECODE => super::dwg_to_dxf_noted(b, Limits { max_dwg_bytes, max_dxf_bytes }),
-        ENCODE => super::dxf_to_dwg_as(b, version).map(|dwg| (dwg, Vec::new())),
+        ENCODE => super::dxf_to_dwg_noted(b, version).map(|(dwg, left_out)| (dwg, left_out.iter().map(|(kind, n)| format!("{kind}\t{n}")).collect())),
         CRASH => std::process::abort(),
         _ => Err(bad()),
     }
@@ -267,6 +283,12 @@ mod tests {
         out
     }
 
+    #[test]
+    fn what_the_writer_left_out_survives_the_reply() {
+        let notes = vec!["ACME_WIDGET\t2".to_string(), "not a count".to_string(), "X\tmany".to_string()];
+        assert_eq!(unwritten_from(&notes), [("ACME_WIDGET".to_string(), 2)].into_iter().collect());
+    }
+
     #[test]
     fn replies_round_trip() {
         let ok: Reply = Ok((b"bytes".to_vec(), vec!["one".into(), "two é".into()]));

crates/engine/src/cmd.rs

@@ -22,7 +22,6 @@ mod inquiry;
 mod layer;
 mod layout;
 mod lisp;
-mod merge;
 mod mleader_content;
 mod mleader_edit;
 mod modify;

crates/engine/src/cmd/blocks/insert_file.rs

@@ -6,7 +6,7 @@
 //! (names this drawing already has keep this drawing's definitions). The file is read through the
 //! host's io hooks like OPEN, so the engine stays I/O-agnostic.
 
-use std::collections::BTreeMap;
+use std::collections::{BTreeMap, BTreeSet};
 use std::sync::Arc;
 
 use cadcraft_doc::{Block, Drawing, Entity, EntityKind, Handle};
@@ -102,7 +102,7 @@ fn define_from(s: &mut Session, src: &Drawing, name: &str) -> Result<String> {
     // Redefining keeps the existing definition's name (block names are case-insensitive).
     let name = d.blocks.keys().find(|k| k.eq_ignore_ascii_case(name)).cloned().unwrap_or_else(|| name.to_string());
     merge_tables(&mut d, src);
-    let renames = crate::cmd::merge::block_renames(&d, src, Some(&name));
+    let renames = block_names(&d, src, &name);
     for (from, to) in &renames {
         let Some(b) = src.blocks.get(from) else { continue };
         let mut nb = (**b).clone();
@@ -136,7 +136,13 @@ fn define_from(s: &mut Session, src: &Drawing, name: &str) -> Result<String> {
 /// Add the source's layers, linetypes and styles that this drawing doesn't have (a name that
 /// exists keeps this drawing's definition).
 fn merge_tables(d: &mut Drawing, src: &Drawing) {
-    use crate::cmd::merge::merge_named as merge;
+    fn merge<T: Clone>(ours: &mut Vec<T>, theirs: &[T], name: fn(&T) -> &str) {
+        for t in theirs {
+            if !ours.iter().any(|o| name(o).eq_ignore_ascii_case(name(t))) {
+                ours.push(t.clone());
+            }
+        }
+    }
     merge(&mut d.layers, &src.layers, |l| &l.name);
     merge(&mut d.linetypes, &src.linetypes, |l| &l.name);
     merge(&mut d.text_styles, &src.text_styles, |t| &t.name);
@@ -146,6 +152,35 @@ fn merge_tables(d: &mut Drawing, src: &Drawing) {
     merge(&mut d.mline_styles, &src.mline_styles, |t| &t.name);
 }
 
+/// Which source blocks to copy, and under which name: new names keep theirs, generated
+/// (anonymous `*…`) blocks that collide get a fresh generated name, and a named block this drawing
+/// already has (or the block being defined) keeps this drawing's definition (not copied).
+fn block_names(d: &Drawing, src: &Drawing, defining: &str) -> BTreeMap<String, String> {
+    let mut out = BTreeMap::new();
+    let mut taken: BTreeSet<String> = d.blocks.keys().map(|k| k.to_ascii_uppercase()).collect();
+    taken.insert(defining.to_ascii_uppercase());
+    for (name, b) in &src.blocks {
+        let upper = name.to_ascii_uppercase();
+        if upper.starts_with("*MODEL_SPACE") || upper.starts_with("*PAPER_SPACE") {
+            continue;
+        }
+        let to = if !taken.contains(&upper) {
+            name.clone()
+        } else if b.anonymous || name.starts_with('*') {
+            let prefix: String = name.chars().take_while(|c| !c.is_ascii_digit()).collect();
+            match (1..=1_000_000).map(|i| format!("{prefix}{i}")).find(|n| !taken.contains(&n.to_ascii_uppercase())) {
+                Some(n) => n,
+                None => continue,
+            }
+        } else {
+            continue;
+        };
+        taken.insert(to.to_ascii_uppercase());
+        out.insert(name.clone(), to);
+    }
+    out
+}
+
 /// A copy of a source entity for this drawing: a fresh handle, block references renamed, and
 /// links to other source objects (associative dimension points) dropped.
 fn adopt(d: &mut Drawing, src: &Drawing, e: &Entity, renames: &BTreeMap<String, String>) -> Entity {

crates/engine/src/cmd/blocks/xref/load.rs

@@ -10,7 +10,6 @@
 //! loaded further up (a circular reference) is not loaded again. Every failure leaves the xref's
 //! block empty with a status saying why; nothing here stops the host drawing from opening.
 
-use std::collections::HashSet;
 use std::path::{Path, PathBuf};
 use std::sync::Arc;
 
@@ -340,15 +339,14 @@ fn load_in(d: &mut Drawing, name: &str, host: Option<&Path>, chosen: bool, cx: &
     let (status, detail) = match fetch(&path, host, chosen, cx, depth) {
         Ok((x, file)) => match import(d, &key, &x, cx) {
             Ok(()) => (XrefStatus::Loaded, file.to_string_lossy().into_owned()),
-            Err(e) => {
-                // Whatever made it in before the failure goes again. A fetch that failed imported
-                // nothing, so it doesn't pay for a second scan of the block table.
-                clear(d, &key);
-                (XrefStatus::Unresolved, e)
-            }
+            Err(e) => (XrefStatus::Unresolved, e),
         },
         Err(e) => e,
     };
+    if status != XrefStatus::Loaded {
+        // Whatever made it in before the failure goes again.
+        clear(d, &key);
+    }
     set_status(d, &key, status);
     (status, detail)
 }
@@ -392,15 +390,9 @@ fn import(d: &mut Drawing, key: &str, x: &Drawing, cx: &mut Ctx) -> Result<(), S
         return Err(format!("more than {MAX_XREF_OBJECTS} objects"));
     }
     let dep = |sym: Sym, n: &str| (!shared(sym, n)).then(|| format!("{key}|{n}"));
-    // The host's names per table, looked up in a set built once: both drawings are hostile input,
-    // so a scan per imported record would be quadratic.
-    fn names<T>(all: &[T], name: fn(&T) -> &str) -> HashSet<String> {
-        all.iter().map(|t| name(t).to_ascii_uppercase()).collect()
-    }
-    let mut have = names(&d.layers, |l| &l.name);
     for l in &x.layers {
         if let Some(n) = dep(Sym::Layer, &l.name)
-            && have.insert(n.to_ascii_uppercase())
+            && d.layer(&n).is_none()
         {
             let mut l = l.clone();
             l.name = n;
@@ -408,44 +400,39 @@ fn import(d: &mut Drawing, key: &str, x: &Drawing, cx: &mut Ctx) -> Result<(), S
             d.layers.push(l);
         }
     }
-    let mut have = names(&d.linetypes, |t| &t.name);
     for t in &x.linetypes {
         if let Some(n) = dep(Sym::Linetype, &t.name)
-            && have.insert(n.to_ascii_uppercase())
+            && d.linetype(&n).is_none()
         {
             d.linetypes.push(cadcraft_doc::Linetype { name: n, ..t.clone() });
         }
     }
-    let mut have = names(&d.text_styles, |t| &t.name);
     for t in &x.text_styles {
         if let Some(n) = dep(Sym::Style, &t.name)
-            && have.insert(n.to_ascii_uppercase())
+            && d.text_style(&n).is_none()
         {
             d.text_styles.push(cadcraft_doc::TextStyle { name: n, ..t.clone() });
         }
     }
-    let mut have = names(&d.dim_styles, |t| &t.name);
     for t in &x.dim_styles {
         if let Some(n) = dep(Sym::DimStyle, &t.name)
-            && have.insert(n.to_ascii_uppercase())
+            && d.dim_style(&n).is_none()
         {
             let text_style = dep(Sym::Style, &t.text_style).unwrap_or_else(|| t.text_style.clone());
             d.dim_styles.push(cadcraft_doc::DimStyle { name: n, text_style, ..t.clone() });
         }
     }
-    let mut have = names(&d.mleader_styles, |t| &t.name);
     for t in &x.mleader_styles {
         if let Some(n) = dep(Sym::MLeaderStyle, &t.name)
-            && have.insert(n.to_ascii_uppercase())
+            && !d.mleader_styles.iter().any(|o| o.name.eq_ignore_ascii_case(&n))
         {
             let text_style = dep(Sym::Style, &t.text_style).unwrap_or_else(|| t.text_style.clone());
             d.mleader_styles.push(cadcraft_doc::MLeaderStyle { name: n, text_style, ..t.clone() });
         }
     }
-    let mut have = names(&d.table_styles, |t| &t.name);
     for t in &x.table_styles {
         if let Some(n) = dep(Sym::TableStyle, &t.name)
-            && have.insert(n.to_ascii_uppercase())
+            && !d.table_styles.iter().any(|o| o.name.eq_ignore_ascii_case(&n))
         {
             d.table_styles.push(cadcraft_doc::TableStyle { name: n, ..t.clone() });
         }

crates/engine/src/cmd/clipboard.rs

@@ -10,7 +10,7 @@
 //! The clipboard can also travel as DXF text ([`system_text`], [`load_system_text`]), so a front
 //! end can put it on the system clipboard and paste it into another CADCraft window.
 
-use std::collections::{BTreeMap, HashSet};
+use std::collections::{BTreeMap, BTreeSet, HashSet};
 use std::sync::Arc;
 
 use cadcraft_doc::{Block, Drawing, Entity, EntityKind, HVal, Handle};
@@ -279,14 +279,20 @@ fn defs_for(src: &Drawing, ents: &[Entity]) -> Drawing {
 /// Add `defs`' table records and blocks that `d` doesn't have. Returns the blocks copied, from
 /// their name in `defs` to their name in `d`.
 fn merge(d: &mut Drawing, defs: &Drawing) -> BTreeMap<String, String> {
-    use crate::cmd::merge::merge_named as add;
+    fn add<T: Clone>(ours: &mut Vec<T>, theirs: &[T], name: fn(&T) -> &str) {
+        for t in theirs {
+            if !ours.iter().any(|o| name(o).eq_ignore_ascii_case(name(t))) {
+                ours.push(t.clone());
+            }
+        }
+    }
     add(&mut d.layers, &defs.layers, |l| &l.name);
     add(&mut d.linetypes, &defs.linetypes, |l| &l.name);
     add(&mut d.text_styles, &defs.text_styles, |t| &t.name);
     add(&mut d.dim_styles, &defs.dim_styles, |t| &t.name);
     add(&mut d.mleader_styles, &defs.mleader_styles, |t| &t.name);
     add(&mut d.table_styles, &defs.table_styles, |t| &t.name);
-    let renames = crate::cmd::merge::block_renames(d, defs, None);
+    let renames = block_names(d, defs);
     for (from, to) in &renames {
         let Some(b) = defs.blocks.get(from) else { continue };
         let mut nb = (**b).clone();
@@ -302,6 +308,34 @@ fn merge(d: &mut Drawing, defs: &Drawing) -> BTreeMap<String, String> {
     renames
 }
 
+/// Which blocks of `defs` to copy, and under which name: a new name keeps its own, a generated
+/// (anonymous `*…`) block whose name is taken gets a fresh generated name, and a named block `d`
+/// already has keeps `d`'s definition (not copied).
+fn block_names(d: &Drawing, defs: &Drawing) -> BTreeMap<String, String> {
+    let mut out = BTreeMap::new();
+    let mut taken: BTreeSet<String> = d.blocks.keys().map(|k| up(k)).collect();
+    for (name, b) in &defs.blocks {
+        let upper = up(name);
+        if upper.starts_with("*MODEL_SPACE") || upper.starts_with("*PAPER_SPACE") {
+            continue;
+        }
+        let to = if !taken.contains(&upper) {
+            name.clone()
+        } else if b.anonymous || name.starts_with('*') {
+            let prefix: String = name.chars().take_while(|c| !c.is_ascii_digit()).collect();
+            match (1..=1_000_000).map(|i| format!("{prefix}{i}")).find(|n| !taken.contains(&up(n))) {
+                Some(n) => n,
+                None => continue,
+            }
+        } else {
+            continue;
+        };
+        taken.insert(up(&to));
+        out.insert(name.clone(), to);
+    }
+    out
+}
+
 /// A copy of a clipboard object for `d`: a fresh handle, block references renamed, and links to
 /// objects of the source drawing (associative dimension points) dropped.
 fn adopt(d: &mut Drawing, e: &Entity, renames: &BTreeMap<String, String>) -> Entity {

crates/engine/src/cmd/exchange.rs

@@ -7,7 +7,7 @@
 //! stays I/O-agnostic; the JSON forms never open dialogs (the UI asks for files when these
 //! commands come from a menu, the tool bar or the command line).
 
-use std::collections::BTreeMap;
+use std::collections::{BTreeMap, BTreeSet};
 use std::sync::Arc;
 
 use cadcraft_color::Rgb;
@@ -166,7 +166,7 @@ fn import_drawing(s: &mut Session, src: &Drawing, at: Vec2, scale: f64, file: &s
     let space = s.space();
     let d = s.doc_mut()?;
     let (layers, styles) = merge_tables(d, src);
-    let renames = crate::cmd::merge::block_renames(d, src, None);
+    let renames = block_names(d, src);
     for (from, to) in &renames {
         let Some(b) = src.blocks.get(from) else { continue };
         let mut nb = (**b).clone();
@@ -213,7 +213,16 @@ fn read_file(_path: &str) -> Result<Vec<u8>> {
 /// Add the source's layers, linetypes and styles that this drawing doesn't have (a name that
 /// exists keeps this drawing's definition). Returns how many layers and how many other records were added.
 fn merge_tables(d: &mut Drawing, src: &Drawing) -> (usize, usize) {
-    use crate::cmd::merge::merge_named as merge;
+    fn merge<T: Clone>(ours: &mut Vec<T>, theirs: &[T], name: fn(&T) -> &str) -> usize {
+        let mut n = 0;
+        for t in theirs {
+            if !ours.iter().any(|o| name(o).eq_ignore_ascii_case(name(t))) {
+                ours.push(t.clone());
+                n += 1;
+            }
+        }
+        n
+    }
     let layers = merge(&mut d.layers, &src.layers, |l| &l.name);
     let others = merge(&mut d.linetypes, &src.linetypes, |l| &l.name)
         + merge(&mut d.text_styles, &src.text_styles, |t| &t.name)
@@ -223,6 +232,35 @@ fn merge_tables(d: &mut Drawing, src: &Drawing) -> (usize, usize) {
     (layers, others)
 }
 
+/// Which source blocks to copy, and under which name: new names keep theirs, generated
+/// (anonymous `*…`) blocks that collide get a fresh generated name, and a named block this drawing
+/// already has keeps this drawing's definition (not copied).
+fn block_names(d: &Drawing, src: &Drawing) -> BTreeMap<String, String> {
+    let mut out = BTreeMap::new();
+    let mut taken: BTreeSet<String> = d.blocks.keys().map(|k| k.to_ascii_uppercase()).collect();
+    for (name, b) in &src.blocks {
+        let upper = name.to_ascii_uppercase();
+        if upper.starts_with("*MODEL_SPACE") || upper.starts_with("*PAPER_SPACE") {
+            continue;
+        }
+        let anonymous = b.anonymous || name.starts_with('*');
+        let to = if !taken.contains(&upper) {
+            name.clone()
+        } else if anonymous {
+            let prefix: String = name.chars().take_while(|c| !c.is_ascii_digit()).collect();
+            match (1..=1_000_000).map(|i| format!("{prefix}{i}")).find(|n| !taken.contains(&n.to_ascii_uppercase())) {
+                Some(n) => n,
+                None => continue,
+            }
+        } else {
+            continue;
+        };
+        taken.insert(to.to_ascii_uppercase());
+        out.insert(name.clone(), to);
+    }
+    out
+}
+
 /// A copy of a source entity for this drawing: a fresh handle, block references renamed, and
 /// links to other source objects (associative dimension points) dropped.
 fn adopt(d: &mut Drawing, e: &Entity, renames: &BTreeMap<String, String>) -> Entity {

crates/engine/src/cmd/file/compare/export.rs

@@ -20,7 +20,7 @@ use std::path::{Path, PathBuf};
 use std::sync::Arc;
 
 use cadcraft_doc::color::{Color, Rgb};
-use cadcraft_doc::{Block, CellStyle, DimStyle, Drawing, Entity, EntityKind, EntityStore, Group, Handle, Layer, MAX_BLOCK_DEPTH, TableStyle};
+use cadcraft_doc::{Block, CellStyle, DimStyle, Drawing, Entity, EntityKind, EntityStore, Group, Handle, MAX_BLOCK_DEPTH, TableStyle};
 use serde_json::{Value, json};
 
 use super::super::super::{CommandSpec, bad};
@@ -197,7 +197,6 @@ fn snapshot(cur: &Drawing, c: &Comparison, folders: &Folders) -> Snapshot {
     d.skipped_records.clear();
     d.handseed = d.handseed.max(other.handseed);
 
-    use crate::cmd::merge::merge_named as merge;
     merge(&mut d.layers, &other.layers, |t| &t.name);
     merge(&mut d.linetypes, &other.linetypes, |t| &t.name);
     merge(&mut d.text_styles, &other.text_styles, |t| &t.name);
@@ -236,17 +235,22 @@ fn snapshot(cur: &Drawing, c: &Comparison, folders: &Folders) -> Snapshot {
     Snapshot { drawing: d, current, compared, counts: (here.len(), there.len()), truncated: diff.truncated, blocks_left_out, uncolored }
 }
 
+/// Add the entries of `from` whose names `to` lacks (compared without letter case).
+fn merge<T: Clone>(to: &mut Vec<T>, from: &[T], name: impl Fn(&T) -> &String) {
+    for t in from {
+        if !to.iter().any(|x| name(x).eq_ignore_ascii_case(name(t))) {
+            to.push(t.clone());
+        }
+    }
+}
+
 /// Add copies of `objs` to model space with new handles, their block references renamed by
 /// `renames` and their files rebased on `folder`, each painted in its comparison color, and
 /// group them as `name` (or `name_2`… when that group name is taken). Returns the group's name.
 fn add_group(d: &mut Drawing, name: &str, objs: &[(&Entity, Color)], renames: &Renames, folder: Option<&Path>, tinted: &mut Tinted) -> String {
     let mut members = Vec::new();
-    // The layers this drawing has, looked up once per object rather than scanned.
-    let mut layers: HashSet<String> = d.layers.iter().map(|l| l.name.to_ascii_uppercase()).collect();
     for (e, color) in objs {
-        if layers.insert(e.common.layer.to_ascii_uppercase()) {
-            d.layers.push(Layer::new(&e.common.layer));
-        }
+        d.ensure_layer(&e.common.layer);
         let h = d.new_handle();
         let mut copy = Entity { handle: h, ..(*e).clone() };
         detach(&mut copy.kind);

crates/engine/src/cmd/merge.rs

@@ -1,99 +0,0 @@
-//! Bringing another drawing's symbol tables and blocks into this one (INSERT of a file, IMPORT,
-//! paste, COMPARE export).
-//!
-//! Both drawings are hostile input, so the work is linear in their sizes: names are looked up in a
-//! set built once per table, and a generated block name continues from the last one allocated for
-//! its prefix instead of probing every taken number again (CWE-407).
-
-use std::collections::{BTreeMap, BTreeSet, HashMap, HashSet};
-
-use cadcraft_doc::Drawing;
-
-/// Add the entries of `theirs` whose names `ours` lacks (compared without letter case; a name that
-/// exists keeps `ours`' definition). Returns how many were added.
-pub(crate) fn merge_named<T: Clone>(ours: &mut Vec<T>, theirs: &[T], name: fn(&T) -> &str) -> usize {
-    let mut have: HashSet<String> = ours.iter().map(|o| name(o).to_ascii_uppercase()).collect();
-    let mut added = 0;
-    for t in theirs {
-        if have.insert(name(t).to_ascii_uppercase()) {
-            ours.push(t.clone());
-            added += 1;
-        }
-    }
-    added
-}
-
-/// Which blocks of `src` to copy into `d`, and under which name: a new name keeps its own, a
-/// generated (anonymous `*…`) block whose name is taken gets a fresh generated name, and a named
-/// block `d` already has (or `defining`, the block being defined from `src`) keeps `d`'s
-/// definition (not copied).
-pub(crate) fn block_renames(d: &Drawing, src: &Drawing, defining: Option<&str>) -> BTreeMap<String, String> {
-    let mut out = BTreeMap::new();
-    let mut taken: BTreeSet<String> = d.blocks.keys().map(|k| k.to_ascii_uppercase()).collect();
-    taken.extend(defining.map(str::to_ascii_uppercase));
-    // The next number to try per generated prefix: numbers below it are taken.
-    let mut next: HashMap<String, usize> = HashMap::new();
-    for (name, b) in &src.blocks {
-        let upper = name.to_ascii_uppercase();
-        if upper.starts_with("*MODEL_SPACE") || upper.starts_with("*PAPER_SPACE") {
-            continue;
-        }
-        let to = if !taken.contains(&upper) {
-            name.clone()
-        } else if b.anonymous || name.starts_with('*') {
-            let prefix: String = name.chars().take_while(|c| !c.is_ascii_digit()).collect();
-            let cursor = next.entry(prefix.to_ascii_uppercase()).or_insert(1);
-            match (*cursor..=1_000_000).map(|i| (i, format!("{prefix}{i}"))).find(|(_, n)| !taken.contains(&n.to_ascii_uppercase())) {
-                Some((i, n)) => {
-                    *cursor = i + 1;
-                    n
-                }
-                None => continue,
-            }
-        } else {
-            continue;
-        };
-        taken.insert(to.to_ascii_uppercase());
-        out.insert(name.clone(), to);
-    }
-    out
-}
-
-#[cfg(test)]
-mod tests {
-    use std::sync::Arc;
-
-    use cadcraft_doc::{Block, Layer};
-
-    use super::*;
-
-    #[test]
-    fn merge_keeps_ours_and_adds_the_rest_without_regard_to_case() {
-        let mut ours = vec![Layer::new("Walls")];
-        let theirs = vec![Layer::new("WALLS"), Layer::new("Doors"), Layer::new("doors")];
-        assert_eq!(merge_named(&mut ours, &theirs, |l| &l.name), 1);
-        assert_eq!(ours.iter().map(|l| l.name.as_str()).collect::<Vec<_>>(), ["Walls", "Doors"]);
-    }
-
-    /// Many colliding anonymous blocks get consecutive names in one pass, not a rescan each.
-    #[test]
-    fn anonymous_collisions_take_consecutive_numbers() {
-        let mut d = Drawing::new_imperial();
-        let mut src = Drawing::new_imperial();
-        for i in 1..=5000 {
-            let mut b = Block::new(&format!("*U{i}"));
-            b.anonymous = true;
-            d.blocks.insert(b.name.clone(), Arc::new(b.clone()));
-            src.blocks.insert(b.name.clone(), Arc::new(b));
-        }
-        let started = std::time::Instant::now();
-        let renames = block_renames(&d, &src, None);
-        assert!(started.elapsed() < std::time::Duration::from_secs(2));
-        assert_eq!(renames.len(), 5000);
-        assert_eq!(renames.get("*U1").map(String::as_str), Some("*U5001"));
-        let distinct: HashSet<&String> = renames.values().collect();
-        assert_eq!(distinct.len(), 5000);
-        // Every new name is the next free number: nothing below 5001 is free, nothing above 10000 is needed.
-        assert!(renames.values().all(|n| n.strip_prefix("*U").and_then(|i| i.parse::<u32>().ok()).is_some_and(|i| (5001..=10000).contains(&i))));
-    }
-}

crates/io/src/lib.rs

@@ -14,6 +14,8 @@ mod dxf_write;
 mod dxf_xref;
 pub mod pdf;
 pub mod plotter;
+/// Open → save → reopen comparison and the real-file corpus harness (gaps FF1, FF5).
+pub mod roundtrip;
 pub mod svg;
 mod svg_read;
 
@@ -69,19 +71,25 @@ fn ext(name: impl AsRef<Path>) -> String {
 /// Read a drawing; the format comes from the content (DXF ASCII/binary, DWG) or the name
 /// (SVG and gzip-compressed SVGZ).
 pub fn read(bytes: &[u8], name: impl AsRef<Path>) -> Result<Drawing> {
-    let name = name.as_ref();
+    read_source(bytes, name.as_ref()).map(|(d, _)| d)
+}
+
+/// [`read`], also returning the DXF rendition a DWG was read through (`None` for other files),
+/// which the round-trip harness counts.
+pub(crate) fn read_source(bytes: &[u8], name: &Path) -> Result<(Drawing, Option<Vec<u8>>)> {
     if cadcraft_dwg::is_dwg(bytes) {
         let (dxf, notes) = cadcraft_dwg::decode(bytes, cadcraft_dwg::Limits::DEFAULT).map_err(IoError::Format)?;
         let mut d = dxf_read::read(&dxf).map_err(|e| IoError::Format(format!("{e} (in the DXF converted from this DWG)")))?;
         d.read_notes = notes;
-        return Ok(d);
+        return Ok((d, Some(dxf)));
     }
-    match (Format::from_path(name), ext(name).as_str()) {
+    let d = match (Format::from_path(name), ext(name).as_str()) {
         (Some(Format::Svg | Format::Svgz), _) => svg_read::read(bytes),
         (Some(Format::Dxf), _) => dxf_read::read(bytes),
         (_, e) if bytes.len() > 4 => dxf_read::read(bytes).map_err(|_| IoError::Unsupported(e.to_string())),
         (_, e) => Err(IoError::Unsupported(e.to_string())),
-    }
+    }?;
+    Ok((d, None))
 }
 
 /// Write a drawing in the format chosen by the name's extension. Images (PNG, SVG, PDF) show
@@ -111,15 +119,29 @@ pub fn write_with(d: &Drawing, name: impl AsRef<Path>, opts: &SaveOptions) -> Re
         // The DWG bridge reads our R2000 DXF and writes the DWG at the version asked for. Without
         // one, at least 2004: R2000 DWG has no true colours, transparency, table styles or
         // gradients, so a drawing from an R2000 DXF doesn't lose them by default.
-        Some(Format::Dwg) if !opts.binary => {
-            let version = opts.version.unwrap_or_else(|| version.max(Version::R2004));
-            cadcraft_dwg::encode_as(dxf_write::dwg_bridge(d, version).as_bytes(), version.code()).map_err(IoError::Format)
-        }
+        Some(Format::Dwg) if !opts.binary => write_dwg(d, opts.version.unwrap_or_else(|| version.max(Version::R2004))).map(|(dwg, _)| dwg),
         _ if opts.binary => Err(IoError::Format(format!("binary output is for DXF files, not `{}`", ext(name)))),
         _ => write_framed(d, name, None),
     }
 }
 
+/// [`write_with`], also returning what the file leaves out of the drawing: for a DWG, how many
+/// records of each DXF type the DWG writer can't write (other programs' custom entities and
+/// objects); for other formats, nothing.
+pub fn write_noted(d: &Drawing, name: impl AsRef<Path>, opts: &SaveOptions) -> Result<(Vec<u8>, cadcraft_dwg::Unwritten)> {
+    let name = name.as_ref();
+    if matches!(Format::from_path(name), Some(Format::Dwg)) && !opts.binary {
+        return write_dwg(d, opts.version.unwrap_or_else(|| save_version(d).max(Version::R2004)));
+    }
+    write_with(d, name, opts).map(|b| (b, cadcraft_dwg::Unwritten::new()))
+}
+
+/// The DWG bridge reads our R2000 DXF and writes the DWG at `version`, in the DWG worker when
+/// there is one.
+fn write_dwg(d: &Drawing, version: Version) -> Result<(Vec<u8>, cadcraft_dwg::Unwritten)> {
+    cadcraft_dwg::encode_noted(dxf_write::dwg_bridge(d, version).as_bytes(), version.code()).map_err(IoError::Format)
+}
+
 /// Like [`write`], but image formats (PNG, SVG, PDF) show `window` (a model-space rectangle in
 /// drawing units, centred and scaled to fit the image without distortion) instead of the
 /// extents. Formats that are not images (DXF, DWG) ignore the window. The window is checked with

crates/io/src/roundtrip/census.rs

@@ -0,0 +1,212 @@
+//! What a DXF file holds, counted from its groups, and the same counts of a drawing read from it:
+//! together they show what opening a file leaves out (the DXF rendition of a DWG for DWG files).
+
+use std::collections::{BTreeMap, HashSet};
+
+use cadcraft_doc::Drawing;
+use serde::Serialize;
+
+use super::{Category, Change, Comparison, Options, Out};
+
+/// Counts of what a file (or a drawing) holds.
+#[derive(Clone, Debug, Default, PartialEq, Eq, Serialize)]
+#[serde(rename_all = "camelCase")]
+pub struct Census {
+    /// Entities of model space and every layout, per DXF type.
+    pub entities: BTreeMap<String, usize>,
+    /// Entities of block definitions (other than the model and paper space blocks), per type.
+    pub block_entities: BTreeMap<String, usize>,
+    /// Block definitions other than the model and paper space blocks and the generated ones.
+    pub blocks: usize,
+    /// Blocks CADCraft generates and so doesn't read (tables' blocks, standard arrowheads
+    /// nothing inserts); their entities are not counted.
+    pub generated_blocks: usize,
+    /// Table records per table type (LAYER, LTYPE, STYLE, DIMSTYLE, VIEW, UCS…).
+    pub tables: BTreeMap<String, usize>,
+    /// Non-graphical objects per type. For a drawing: only those kept without being modelled.
+    pub objects: BTreeMap<String, usize>,
+    /// CLASS records. For a drawing: only those kept for unmodelled objects.
+    pub classes: usize,
+}
+
+/// Records that belong to the record before them (or delimit blocks), not entities of their own.
+const SUB_RECORDS: &[&str] = &["VERTEX", "SEQEND", "ATTRIB", "BLOCK", "ENDBLK"];
+
+fn bump(m: &mut BTreeMap<String, usize>, k: &str) {
+    let n = m.entry(k.to_string()).or_insert(0);
+    *n = n.saturating_add(1);
+}
+
+/// A layout's own paper-space viewport (VIEWPORT with ID 1, group 69): part of the layout, not
+/// an entity of its own for CADCraft.
+fn is_overall_viewport(t: &str, g: &[cadcraft_dxf::Tag]) -> bool {
+    t == "VIEWPORT" && g.iter().any(|x| x.code == 69 && x.i64() == 1)
+}
+
+fn is_space_block(name: &str) -> bool {
+    let up = name.trim().to_ascii_uppercase();
+    up.starts_with("*MODEL_SPACE") || up.starts_with("*PAPER_SPACE")
+}
+
+/// Count what a DXF file (ASCII or binary) holds; `None` when it can't be parsed.
+pub fn census(dxf: &[u8]) -> Option<Census> {
+    let tags = cadcraft_dxf::parse(dxf).ok()?;
+    let mut c = Census::default();
+    // Blocks CADCraft generates (tables' blocks, standard arrowheads nothing inserts) are left out
+    // on read and rebuilt on save: counted apart.
+    let mut table_blocks: HashSet<String> = HashSet::new();
+    let mut inserted: HashSet<String> = HashSet::new();
+    let mut blocks: Vec<(String, BTreeMap<String, usize>)> = Vec::new();
+    let mut note_refs = |t: &str, g: &[cadcraft_dxf::Tag]| {
+        let name = || g.iter().find(|x| x.code == 2).map(|x| x.str().trim().to_ascii_uppercase());
+        match t {
+            "ACAD_TABLE" => table_blocks.extend(name()),
+            "INSERT" => inserted.extend(name()),
+            _ => {}
+        }
+    };
+    for s in cadcraft_dxf::sections(&tags) {
+        let records = cadcraft_dxf::records(&s.tags);
+        match s.name.trim().to_ascii_uppercase().as_str() {
+            "ENTITIES" => {
+                for (t, g) in &records {
+                    note_refs(t, g);
+                    if !SUB_RECORDS.contains(&t.as_str()) && !is_overall_viewport(t, g) {
+                        bump(&mut c.entities, t);
+                    }
+                }
+            }
+            "BLOCKS" => {
+                let mut space = false;
+                for (t, g) in &records {
+                    note_refs(t, g);
+                    match t.as_str() {
+                        "BLOCK" => {
+                            let name = g.iter().find(|x| x.code == 2).map(cadcraft_dxf::Tag::str).unwrap_or_default();
+                            space = is_space_block(&name);
+                            if !space {
+                                blocks.push((name.trim().to_ascii_uppercase(), BTreeMap::new()));
+                            }
+                        }
+                        t if SUB_RECORDS.contains(&t) || is_overall_viewport(t, g) => {}
+                        t if space => bump(&mut c.entities, t),
+                        t => {
+                            if let Some((_, m)) = blocks.last_mut() {
+                                bump(m, t);
+                            }
+                        }
+                    }
+                }
+            }
+            "TABLES" => {
+                for (t, _) in &records {
+                    if t != "TABLE" && t != "ENDTAB" {
+                        bump(&mut c.tables, t);
+                    }
+                }
+            }
+            "OBJECTS" => {
+                for (t, _) in &records {
+                    bump(&mut c.objects, t);
+                }
+            }
+            "CLASSES" => c.classes = c.classes.saturating_add(records.iter().filter(|(t, _)| t == "CLASS").count()),
+            _ => {}
+        }
+    }
+    for (name, m) in blocks {
+        if table_blocks.contains(&name) || (crate::dxf_ext::is_standard_arrow(&name) && !inserted.contains(&name)) {
+            c.generated_blocks = c.generated_blocks.saturating_add(1);
+            continue;
+        }
+        c.blocks = c.blocks.saturating_add(1);
+        for (t, n) in m {
+            let slot = c.block_entities.entry(t).or_insert(0);
+            *slot = slot.saturating_add(n);
+        }
+    }
+    Some(c)
+}
+
+/// The same counts of a drawing: its entities, blocks and table records, and the objects and
+/// classes it keeps without modelling them.
+pub fn census_of(d: &Drawing) -> Census {
+    let mut c = Census::default();
+    for e in d.model.iter().chain(d.layouts.iter().flat_map(|l| l.entities.iter())) {
+        bump(&mut c.entities, e.kind.dxf_name());
+    }
+    for b in d.blocks.values() {
+        for e in b.entities.iter() {
+            bump(&mut c.block_entities, e.kind.dxf_name());
+        }
+    }
+    c.blocks = d.blocks.len();
+    for (t, n) in [
+        ("LAYER", d.layers.len()),
+        ("LTYPE", d.linetypes.len()),
+        ("STYLE", d.text_styles.len()),
+        ("DIMSTYLE", d.dim_styles.len()),
+        ("VIEW", d.views.len()),
+        ("UCS", d.ucss.len()),
+    ] {
+        c.tables.insert(t.to_string(), n);
+    }
+    for o in &d.foreign.objects {
+        bump(&mut c.objects, &o.kind);
+    }
+    c.classes = d.foreign.classes.len();
+    c
+}
+
+/// Polylines may change between the old-style POLYLINE and LWPOLYLINE when read: counted
+/// together.
+fn merged(m: &BTreeMap<String, usize>) -> BTreeMap<String, usize> {
+    let mut out = BTreeMap::new();
+    for (k, n) in m {
+        let key = if k == "POLYLINE" || k == "LWPOLYLINE" { "POLYLINE+LWPOLYLINE" } else { k.as_str() };
+        let slot = out.entry(key.to_string()).or_insert(0usize);
+        *slot = slot.saturating_add(*n);
+    }
+    out
+}
+
+fn compare_counts(out: &mut Out, place: &str, file: &BTreeMap<String, usize>, read: &BTreeMap<String, usize>) {
+    let (file, read) = (merged(file), merged(read));
+    let keys: std::collections::BTreeSet<&String> = file.keys().chain(read.keys()).collect();
+    for k in keys {
+        let (f, r) = (file.get(k).copied().unwrap_or(0), read.get(k).copied().unwrap_or(0));
+        if r < f {
+            out.push_n(Category::Read, Change::Lost, f - r, format!("{place} {k}"), format!("{k}: {f} in the file, {r} read"));
+        } else if r > f {
+            out.push_n(Category::Read, Change::Added, r - f, format!("{place} {k}"), format!("{k}: {f} in the file, {r} read"));
+        }
+    }
+}
+
+/// What a drawing read from a file holds less (or more) than the file: entities per type in the
+/// spaces and in blocks, blocks, and the table records CADCraft models. Blocks CADCraft generates
+/// (tables' blocks, unused standard arrowheads) are left out on both sides.
+pub fn compare_census(file: &Census, d: &Drawing, opts: &Options) -> Comparison {
+    let read = census_of(d);
+    let mut out = Out::new(opts);
+    compare_counts(&mut out, "spaces", &file.entities, &read.entities);
+    compare_counts(&mut out, "blocks", &file.block_entities, &read.block_entities);
+    if file.blocks != read.blocks {
+        let change = if read.blocks < file.blocks { Change::Lost } else { Change::Added };
+        out.push_n(
+            Category::Read,
+            change,
+            file.blocks.abs_diff(read.blocks),
+            "BLOCK definitions",
+            format!("{} in the file, {} read", file.blocks, read.blocks),
+        );
+    }
+    for t in ["LAYER", "LTYPE", "STYLE", "DIMSTYLE", "VIEW", "UCS"] {
+        let (f, r) = (file.tables.get(t).copied().unwrap_or(0), read.tables.get(t).copied().unwrap_or(0));
+        if f != r {
+            let change = if r < f { Change::Lost } else { Change::Added };
+            out.push_n(Category::Read, change, f.abs_diff(r), format!("{t} table"), format!("{f} records in the file, {r} read"));
+        }
+    }
+    out.cmp
+}

crates/io/src/roundtrip/corpus.rs

@@ -0,0 +1,695 @@
+//! The corpus harness: every DXF and DWG file of a folder is opened, saved as DXF and as DWG,
+//! reopened and compared with the drawing first opened; the findings go to a Markdown and a JSON
+//! report. A file that fails, or even makes CADCraft panic, is a reported failure: each stage runs
+//! under `catch_unwind` here, as a last resort of the harness only.
+
+use std::collections::BTreeMap;
+use std::panic::AssertUnwindSafe;
+use std::path::{Path, PathBuf};
+use std::time::Instant;
+
+use cadcraft_doc::Drawing;
+use serde::Serialize;
+
+use super::{Category, Census, Change, Comparison, Options, Tally, census, census_of, clip, compare, compare_census};
+
+/// A drawing file format the harness checks.
+#[derive(Clone, Copy, Debug, PartialEq, Eq, Serialize)]
+#[serde(rename_all = "lowercase")]
+pub enum Format {
+    Dxf,
+    Dwg,
+}
+
+impl Format {
+    pub fn ext(self) -> &'static str {
+        match self {
+            Format::Dxf => "dxf",
+            Format::Dwg => "dwg",
+        }
+    }
+    pub fn label(self) -> &'static str {
+        match self {
+            Format::Dxf => "DXF",
+            Format::Dwg => "DWG",
+        }
+    }
+    /// The format of a file's bytes: DWG by its magic, otherwise DXF.
+    pub fn of(bytes: &[u8]) -> Format {
+        if cadcraft_dwg::is_dwg(bytes) { Format::Dwg } else { Format::Dxf }
+    }
+    fn other(self) -> Format {
+        match self {
+            Format::Dxf => Format::Dwg,
+            Format::Dwg => Format::Dxf,
+        }
+    }
+}
+
+/// How the harness runs.
+#[derive(Clone, Debug, PartialEq)]
+pub struct CorpusOptions {
+    pub compare: Options,
+    /// Also save in the other format (DWG → DXF, DXF → DWG).
+    pub cross_format: bool,
+    /// Count what each file holds, to report what opening it leaves out.
+    pub census: bool,
+    /// Keep the saved files in this folder (and reopen them from there).
+    pub keep_saved: Option<PathBuf>,
+}
+
+impl Default for CorpusOptions {
+    fn default() -> Self {
+        CorpusOptions { compare: Options::default(), cross_format: true, census: true, keep_saved: None }
+    }
+}
+
+/// Largest file the harness opens.
+pub const MAX_FILE_BYTES: u64 = 1 << 30;
+/// Deepest folder nesting walked.
+const MAX_DEPTH: usize = 32;
+/// Most files taken from one corpus.
+const MAX_FILES: usize = 100_000;
+
+/// One save → reopen of a drawing.
+#[derive(Clone, Debug, Default, PartialEq, Serialize)]
+#[serde(rename_all = "camelCase")]
+pub struct RouteReport {
+    /// `dwg → dxf`.
+    pub route: String,
+    pub saved_bytes: usize,
+    /// Where the saved file was kept, when it was.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub saved_path: Option<String>,
+    pub save_error: Option<String>,
+    pub reopen_error: Option<String>,
+    /// What reopening reported (records the reader couldn't place or decode).
+    pub reopen_notes: Vec<String>,
+    /// What the writer itself left out of the saved file: records per DXF type (upper case) it
+    /// can't write in that format (for DWG: other programs' custom entities and objects). Each
+    /// is lost; the reopened drawing's entity counts show the same entities missing, and a strict
+    /// corpus run fails on any.
+    pub not_written: BTreeMap<String, usize>,
+    /// The reopened drawing against the drawing first opened.
+    pub comparison: Comparison,
+    pub seconds: f64,
+}
+
+impl RouteReport {
+    pub fn ok(&self) -> bool {
+        self.save_error.is_none() && self.reopen_error.is_none()
+    }
+    /// Saved and reopened with no difference at all.
+    pub fn lossless(&self) -> bool {
+        self.ok() && self.not_written.is_empty() && self.comparison.is_empty()
+    }
+    /// How many records the writer left out ([`Self::not_written`]).
+    pub fn not_written_total(&self) -> usize {
+        self.not_written.values().fold(0usize, |a, n| a.saturating_add(*n))
+    }
+}
+
+/// `2 ACME_WIDGET, 1 OLE2FRAME`.
+fn per_type(m: &BTreeMap<String, usize>) -> String {
+    m.iter().map(|(t, n)| format!("{n} {t}")).collect::<Vec<_>>().join(", ")
+}
+
+/// Everything found for one file.
+#[derive(Clone, Debug, Default, PartialEq, Serialize)]
+#[serde(rename_all = "camelCase")]
+pub struct FileReport {
+    /// The path relative to the corpus folder.
+    pub path: String,
+    pub format: Option<Format>,
+    /// `AC1032` for a DWG, the `$ACADVER` of a DXF.
+    pub version: String,
+    pub bytes: u64,
+    pub open_error: Option<String>,
+    /// What opening reported (a DWG's records the reader skipped, damaged records).
+    pub open_notes: Vec<String>,
+    pub open_seconds: f64,
+    /// What the file holds (for a DWG: its DXF rendition), when counted.
+    pub file_census: Option<Census>,
+    /// What the drawing read from it holds.
+    pub drawing_census: Option<Census>,
+    /// What the drawing holds less or more than the file.
+    pub read: Comparison,
+    pub routes: Vec<RouteReport>,
+}
+
+impl FileReport {
+    pub fn opened(&self) -> bool {
+        self.open_error.is_none()
+    }
+}
+
+impl CorpusReport {
+    /// What makes a strict corpus run fail, one line each: a file that didn't open, a round trip
+    /// that didn't save or reopen, and records a writer left out ([`RouteReport::not_written`]).
+    /// Other differences are reported, not failed on.
+    pub fn strict_failures(&self) -> Vec<String> {
+        let mut out = Vec::new();
+        for f in &self.files {
+            if !f.opened() {
+                out.push(format!("{}: failed to open", f.path));
+            }
+            for r in &f.routes {
+                if !r.ok() {
+                    out.push(format!("{} ({}): failed to save or reopen", f.path, r.route));
+                }
+                if !r.not_written.is_empty() {
+                    out.push(format!("{} ({}): lost on save, not written: {}", f.path, r.route, per_type(&r.not_written)));
+                }
+            }
+        }
+        out
+    }
+}
+
+/// The whole corpus run.
+#[derive(Clone, Debug, Default, PartialEq, Serialize)]
+#[serde(rename_all = "camelCase")]
+pub struct CorpusReport {
+    pub corpus: String,
+    /// UTC, `2026-10-21 09:30`.
+    pub generated: String,
+    pub cadcraft_version: String,
+    pub abs_tol: f64,
+    pub rel_tol: f64,
+    pub files: Vec<FileReport>,
+}
+
+/// Run `f`, turning a panic into an error message.
+fn guard<T>(f: impl FnOnce() -> T) -> Result<T, String> {
+    std::panic::catch_unwind(AssertUnwindSafe(f)).map_err(|p| {
+        let msg =
+            p.downcast_ref::<&str>().map(|s| s.to_string()).or_else(|| p.downcast_ref::<String>().cloned()).unwrap_or_else(|| "no message".into());
+        format!("CRASH (panic): {msg}")
+    })
+}
+
+/// A drawing's notes on what reading it left out.
+fn notes(d: &Drawing) -> Vec<String> {
+    let mut v = d.read_notes.clone();
+    for (section, n) in &d.skipped_records {
+        v.push(format!("{n} record(s) of the {section} section couldn't be placed and were skipped"));
+    }
+    v
+}
+
+/// Save `d` as `to`, reopen it and compare. `keep` is the file to keep the saved bytes in (and
+/// reopen them from).
+pub fn round_trip(d: &Drawing, from: Format, to: Format, opts: &Options, keep: Option<&Path>) -> RouteReport {
+    let start = Instant::now();
+    let mut r = RouteReport { route: format!("{} → {}", from.ext(), to.ext()), ..RouteReport::default() };
+    let name = format!("roundtrip.{}", to.ext());
+    let saved = match guard(|| crate::write_noted(d, &name, &crate::SaveOptions::default())) {
+        Ok(Ok((b, not_written))) => {
+            r.not_written = not_written;
+            b
+        }
+        Ok(Err(e)) => {
+            r.save_error = Some(e.to_string());
+            r.seconds = start.elapsed().as_secs_f64();
+            return r;
+        }
+        Err(e) => {
+            r.save_error = Some(e);
+            r.seconds = start.elapsed().as_secs_f64();
+            return r;
+        }
+    };
+    r.saved_bytes = saved.len();
+    let mut bytes = saved;
+    if let Some(path) = keep {
+        let kept = path.parent().map_or(Ok(()), std::fs::create_dir_all).and_then(|_| std::fs::write(path, &bytes)).and_then(|_| std::fs::read(path));
+        match kept {
+            Ok(b) => {
+                bytes = b;
+                r.saved_path = Some(path.display().to_string());
+            }
+            Err(e) => r.reopen_notes.push(format!("couldn't keep the saved file at {}: {e}", path.display())),
+        }
+    }
+    let reopened = match guard(|| crate::read(&bytes, &name)) {
+        Ok(Ok(d2)) => d2,
+        Ok(Err(e)) => {
+            r.reopen_error = Some(e.to_string());
+            r.seconds = start.elapsed().as_secs_f64();
+            return r;
+        }
+        Err(e) => {
+            r.reopen_error = Some(e);
+            r.seconds = start.elapsed().as_secs_f64();
+            return r;
+        }
+    };
+    r.reopen_notes.extend(notes(&reopened));
+    match guard(|| compare(d, &reopened, opts)) {
+        Ok(c) => r.comparison = c,
+        Err(e) => r.reopen_error = Some(format!("comparing: {e}")),
+    }
+    attribute_unwritten(&mut r, to);
+    r.seconds = start.elapsed().as_secs_f64();
+    r
+}
+
+/// Say on the lost entity rows of a type the writer couldn't write that this is why they are
+/// lost, so a writer's gap reads as one and isn't taken for a reader's or a comparison's.
+fn attribute_unwritten(r: &mut RouteReport, to: Format) {
+    if r.not_written.is_empty() {
+        return;
+    }
+    for d in r.comparison.differences.iter_mut().filter(|d| d.category == Category::Count && d.change == Change::Lost) {
+        if let Some(t) = r.not_written.keys().find(|t| d.detail.starts_with(&format!("{t}: "))) {
+            d.detail.push_str(&format!(" (not written: the {} writer can't write {t})", to.label()));
+        }
+    }
+}
+
+/// The kept saved file of `rel` saved as `to`, in `dir`.
+fn kept_path(dir: &Path, rel: &str, to: Format) -> PathBuf {
+    let flat: String = rel.chars().map(|c| if matches!(c, '/' | '\\' | ':') { '_' } else { c }).collect();
+    dir.join(format!("{flat}.roundtrip.{}", to.ext()))
+}
+
+/// Check one file's bytes: open, count, save in each format, reopen, compare. `rel` names it in
+/// the report.
+pub fn check_bytes(bytes: &[u8], rel: &str, opts: &CorpusOptions) -> FileReport {
+    let format = Format::of(bytes);
+    let mut f = FileReport { path: rel.to_string(), format: Some(format), bytes: bytes.len() as u64, ..FileReport::default() };
+    if format == Format::Dwg {
+        f.version = cadcraft_dwg::version(bytes).unwrap_or_default();
+    }
+    let start = Instant::now();
+    let name = format!("source.{}", format.ext());
+    let opened = guard(|| crate::read_source(bytes, std::path::Path::new(&name)));
+    f.open_seconds = start.elapsed().as_secs_f64();
+    let (d, dxf) = match opened {
+        Ok(Ok(x)) => x,
+        Ok(Err(e)) => {
+            f.open_error = Some(e.to_string());
+            return f;
+        }
+        Err(e) => {
+            f.open_error = Some(e);
+            return f;
+        }
+    };
+    if format == Format::Dxf {
+        f.version = d.header.str("ACADVER", "");
+    }
+    f.open_notes = notes(&d);
+    f.drawing_census = guard(|| census_of(&d)).ok();
+    if opts.census {
+        let source = dxf.as_deref().unwrap_or(bytes);
+        match guard(|| census(source)) {
+            Ok(Some(c)) => {
+                if let Ok(read) = guard(|| compare_census(&c, &d, &opts.compare)) {
+                    f.read = read;
+                }
+                f.file_census = Some(c);
+            }
+            Ok(None) => f.open_notes.push("the file's groups couldn't be counted".into()),
+            Err(e) => f.open_notes.push(format!("counting the file's groups: {e}")),
+        }
+    }
+    let mut targets = vec![format];
+    if opts.cross_format {
+        targets.push(format.other());
+    }
+    for to in targets {
+        let keep = opts.keep_saved.as_deref().map(|dir| kept_path(dir, rel, to));
+        f.routes.push(round_trip(&d, format, to, &opts.compare, keep.as_deref()));
+    }
+    f
+}
+
+/// Check one file on disk; `rel` names it in the report.
+pub fn check_file(path: &Path, rel: &str, opts: &CorpusOptions) -> FileReport {
+    let fail = |e: String| FileReport { path: rel.to_string(), open_error: Some(e), ..FileReport::default() };
+    match std::fs::metadata(path) {
+        Ok(m) if m.len() > MAX_FILE_BYTES => return fail(format!("the file is {} bytes, more than the {MAX_FILE_BYTES} the harness opens", m.len())),
+        Ok(_) => {}
+        Err(e) => return fail(format!("can't read the file: {e}")),
+    }
+    match std::fs::read(path) {
+        Ok(bytes) => check_bytes(&bytes, rel, opts),
+        Err(e) => fail(format!("can't read the file: {e}")),
+    }
+}
+
+/// Whether a path names a DXF or DWG file (by extension).
+pub fn is_drawing_file(path: &Path) -> bool {
+    path.extension().map(|e| e.to_string_lossy().to_ascii_lowercase()).is_some_and(|e| e == "dxf" || e == "dwg")
+}
+
+/// Every DXF and DWG file under `dir`, sorted, leaving out hidden folders and the folders in
+/// `skip` (the report folder). Symbolic links are not followed.
+pub fn corpus_files(dir: &Path, skip: &[PathBuf]) -> Vec<PathBuf> {
+    walk(dir, skip).files
+}
+
+/// What [`walk`] found, and what it couldn't look at, so a report never passes for a whole one.
+#[derive(Debug, Default)]
+pub struct Walk {
+    pub files: Vec<PathBuf>,
+    /// Folders and entries that couldn't be read, with why.
+    pub unreadable: Vec<String>,
+    /// Files past [`MAX_FILES`] or folders past the depth limit were left out.
+    pub truncated: bool,
+}
+
+/// [`corpus_files`], also saying what it had to leave out.
+pub fn walk(dir: &Path, skip: &[PathBuf]) -> Walk {
+    let skip: Vec<PathBuf> = skip.iter().map(|p| p.canonicalize().unwrap_or_else(|_| p.clone())).collect();
+    let mut w = Walk::default();
+    let mut stack = vec![(dir.to_path_buf(), 0usize)];
+    while let Some((d, depth)) = stack.pop() {
+        let entries = match std::fs::read_dir(&d) {
+            Ok(entries) => entries,
+            Err(e) => {
+                w.unreadable.push(format!("{}: {e}", d.display()));
+                continue;
+            }
+        };
+        for entry in entries {
+            let entry = match entry {
+                Ok(entry) => entry,
+                Err(e) => {
+                    w.unreadable.push(format!("{}: {e}", d.display()));
+                    continue;
+                }
+            };
+            let path = entry.path();
+            let ft = match entry.file_type() {
+                Ok(ft) => ft,
+                Err(e) => {
+                    w.unreadable.push(format!("{}: {e}", path.display()));
+                    continue;
+                }
+            };
+            if ft.is_dir() {
+                let hidden = entry.file_name().to_string_lossy().starts_with('.');
+                let canon = path.canonicalize().unwrap_or_else(|_| path.clone());
+                if hidden || skip.contains(&canon) {
+                    continue;
+                }
+                if depth < MAX_DEPTH {
+                    stack.push((path, depth + 1));
+                } else {
+                    w.truncated = true;
+                }
+            } else if ft.is_file() && is_drawing_file(&path) {
+                if w.files.len() < MAX_FILES {
+                    w.files.push(path);
+                } else {
+                    w.truncated = true;
+                }
+            }
+        }
+    }
+    w.files.sort();
+    w.unreadable.sort();
+    w
+}
+
+/// `path` relative to `root`, with forward slashes.
+pub fn relative(path: &Path, root: &Path) -> String {
+    let rel = path.strip_prefix(root).unwrap_or(path);
+    rel.components().map(|c| c.as_os_str().to_string_lossy().into_owned()).collect::<Vec<_>>().join("/")
+}
+
+/// Check `files` (paths under `root`, which names the corpus in the report), calling `progress`
+/// with (index, count, relative path) before each.
+pub fn run(root: &Path, files: &[PathBuf], opts: &CorpusOptions, progress: &mut dyn FnMut(usize, usize, &str)) -> CorpusReport {
+    let mut report = CorpusReport {
+        corpus: root.display().to_string(),
+        generated: now_utc(),
+        cadcraft_version: env!("CARGO_PKG_VERSION").to_string(),
+        abs_tol: opts.compare.abs_tol,
+        rel_tol: opts.compare.rel_tol,
+        files: Vec::new(),
+    };
+    for (i, path) in files.iter().enumerate() {
+        let rel = relative(path, root);
+        progress(i, files.len(), &rel);
+        let f = match guard(|| check_file(path, &rel, opts)) {
+            Ok(f) => f,
+            Err(e) => FileReport { path: rel, open_error: Some(e), ..FileReport::default() },
+        };
+        report.files.push(f);
+    }
+    report
+}
+
+/// The current time, UTC, as `2026-10-21 09:30`.
+fn now_utc() -> String {
+    let secs = std::time::SystemTime::now().duration_since(std::time::UNIX_EPOCH).map(|d| d.as_secs()).unwrap_or(0);
+    let days = i64::try_from(secs / 86_400).unwrap_or(0);
+    let (h, m) = ((secs % 86_400) / 3600, (secs % 3600) / 60);
+    // Days to a civil date (proleptic Gregorian), after Howard Hinnant's public-domain algorithm.
+    let z = days + 719_468;
+    let era = z.div_euclid(146_097);
+    let doe = z.rem_euclid(146_097);
+    let yoe = (doe - doe / 1460 + doe / 36_524 - doe / 146_096) / 365;
+    let doy = doe - (365 * yoe + yoe / 4 - yoe / 100);
+    let mp = (5 * doy + 2) / 153;
+    let day = doy - (153 * mp + 2) / 5 + 1;
+    let month = if mp < 10 { mp + 3 } else { mp - 9 };
+    let year = yoe + era * 400 + i64::from(month <= 2);
+    format!("{year:04}-{month:02}-{day:02} {h:02}:{m:02} UTC")
+}
+
+// ---------------------------------------------------------------------------------------------
+// Rendering
+
+/// Most differences listed per round trip in the Markdown report (all are in the JSON).
+const MD_ROWS: usize = 300;
+
+/// Text for a Markdown table cell.
+fn cell(s: &str) -> String {
+    s.replace('|', "\\|").replace(['\r', '\n'], " ")
+}
+
+fn size(n: u64) -> String {
+    if n >= 1 << 20 {
+        format!("{:.1} MB", n as f64 / f64::from(1u32 << 20))
+    } else if n >= 1 << 10 {
+        format!("{:.0} KB", n as f64 / 1024.0)
+    } else {
+        format!("{n} B")
+    }
+}
+
+fn status(err: &Option<String>) -> String {
+    match err {
+        None => "ok".into(),
+        Some(e) if e.starts_with("CRASH") => "**CRASH**".into(),
+        Some(_) => "**failed**".into(),
+    }
+}
+
+fn tally_cells(t: Tally) -> String {
+    format!("{} | {} | {}", t.lost, t.added, t.changed)
+}
+
+impl CorpusReport {
+    pub fn to_json(&self) -> String {
+        serde_json::to_string_pretty(self).unwrap_or_else(|e| format!("{{\"error\": \"{e}\"}}"))
+    }
+
+    /// Totals per category over every round trip.
+    pub fn totals(&self) -> std::collections::BTreeMap<Category, Tally> {
+        let mut all: std::collections::BTreeMap<Category, Tally> = std::collections::BTreeMap::new();
+        for r in self.files.iter().flat_map(|f| f.routes.iter()) {
+            for (c, t) in &r.comparison.totals {
+                let slot = all.entry(*c).or_default();
+                slot.add(Change::Lost, t.lost);
+                slot.add(Change::Added, t.added);
+                slot.add(Change::Changed, t.changed);
+            }
+        }
+        all
+    }
+
+    pub fn to_markdown(&self) -> String {
+        let mut s = String::new();
+        let opened = self.files.iter().filter(|f| f.opened()).count();
+        let routes: Vec<&RouteReport> = self.files.iter().flat_map(|f| f.routes.iter()).collect();
+        let save_failed = routes.iter().filter(|r| r.save_error.is_some()).count();
+        let reopen_failed = routes.iter().filter(|r| r.reopen_error.is_some()).count();
+        let lossless = routes.iter().filter(|r| r.lossless()).count();
+        let crashes = self.files.iter().filter(|f| f.open_error.as_deref().is_some_and(|e| e.starts_with("CRASH"))).count()
+            + routes.iter().filter(|r| [&r.save_error, &r.reopen_error].iter().any(|e| e.as_deref().is_some_and(|e| e.contains("CRASH")))).count();
+        s.push_str("# CADCraft round-trip report\n\n");
+        s.push_str(&format!("- Corpus: `{}`\n", self.corpus));
+        s.push_str(&format!("- Generated: {} by CADCraft {}\n", self.generated, self.cadcraft_version));
+        s.push_str(&format!("- Tolerance for numbers: {:e} absolute + {:e} relative\n", self.abs_tol, self.rel_tol));
+        s.push_str(&format!(
+            "- Files: {} ({opened} opened, {} failed to open); round trips: {} ({save_failed} failed to save, {reopen_failed} failed to reopen, {lossless} without any difference); crashes: {crashes}\n",
+            self.files.len(),
+            self.files.len() - opened,
+            routes.len()
+        ));
+        let mut unwritten: BTreeMap<String, usize> = BTreeMap::new();
+        for r in &routes {
+            for (t, n) in &r.not_written {
+                let slot = unwritten.entry(format!("{t} ({})", r.route)).or_insert(0);
+                *slot = slot.saturating_add(*n);
+            }
+        }
+        if unwritten.is_empty() {
+            s.push_str("- Lost on save (records a writer can't write): none\n\n");
+        } else {
+            s.push_str(&format!("- **Lost on save** (records a writer can't write): {}\n\n", per_type(&unwritten)));
+        }
+        s.push_str(
+            "Each file is opened, saved in its own format and in the other one, and the saved file is reopened and compared with \
+             the drawing first opened. *Read* counts what the file holds that the drawing read from it doesn't (for a DWG: its \
+             DXF rendition). Lost/added/changed count items (entities, records, values).\n\n",
+        );
+        s.push_str("## Summary\n\n");
+        s.push_str("| File | Format | Size | Open | Read lost | Round trip | Save | Reopen | Not written | Lost | Added | Changed |\n");
+        s.push_str("|---|---|---|---|---|---|---|---|---|---|---|---|\n");
+        for f in &self.files {
+            let fmt = f.format.map_or("?".to_string(), |x| format!("{} {}", x.label(), f.version).trim().to_string());
+            let open = if f.opened() { format!("ok ({:.1} s)", f.open_seconds) } else { status(&f.open_error) };
+            let lead = format!("| `{}` | {} | {} | {} | {} ", cell(&f.path), cell(&fmt), size(f.bytes), open, f.read.tally(Category::Read).lost);
+            if f.routes.is_empty() {
+                s.push_str(&format!("{lead}| — | — | — | — | — | — | — |\n"));
+            }
+            for (i, r) in f.routes.iter().enumerate() {
+                let head = if i == 0 { lead.clone() } else { "| | | | | ".to_string() };
+                let t = r.comparison.total();
+                let reopen = if r.save_error.is_some() { "—".to_string() } else { status(&r.reopen_error) };
+                let unwritten = if r.not_written.is_empty() { "0".to_string() } else { format!("**{}**", r.not_written_total()) };
+                s.push_str(&format!("{head}| {} | {} | {} | {unwritten} | {} |\n", r.route, status(&r.save_error), reopen, tally_cells(t)));
+            }
+        }
+        s.push_str("\n## By category (all round trips)\n\n| Category | Lost | Added | Changed |\n|---|---|---|---|\n");
+        let totals = self.totals();
+        for c in Category::ALL {
+            if let Some(t) = totals.get(&c).filter(|t| t.total() > 0) {
+                s.push_str(&format!("| {} | {} |\n", c.label(), tally_cells(*t)));
+            }
+        }
+        if totals.values().all(|t| t.total() == 0) {
+            s.push_str("| (no differences) | 0 | 0 | 0 |\n");
+        }
+        s.push_str("\n## Files\n");
+        for f in &self.files {
+            self.file_markdown(f, &mut s);
+        }
+        s
+    }
+
+    fn file_markdown(&self, f: &FileReport, s: &mut String) {
+        s.push_str(&format!("\n### `{}`\n\n", f.path));
+        let fmt = f.format.map_or("unknown format".to_string(), |x| format!("{} {}", x.label(), f.version).trim().to_string());
+        s.push_str(&format!("- {fmt}, {}\n", size(f.bytes)));
+        match &f.open_error {
+            Some(e) => {
+                s.push_str(&format!("- Open: {}: {}\n", status(&f.open_error), cell(e)));
+                return;
+            }
+            None => {
+                let what = f.drawing_census.as_ref().map_or(String::new(), |c| {
+                    let n: usize = c.entities.values().sum();
+                    let nb: usize = c.block_entities.values().sum();
+                    format!(
+                        ": {n} entities in model and paper space, {} blocks holding {nb} entities, {} layers, {} unmodelled objects kept",
+                        c.blocks,
+                        c.tables.get("LAYER").copied().unwrap_or(0),
+                        c.objects.values().sum::<usize>()
+                    )
+                });
+                s.push_str(&format!("- Open: ok in {:.1} s{what}\n", f.open_seconds));
+            }
+        }
+        for n in &f.open_notes {
+            s.push_str(&format!("- Note on opening: {}\n", cell(n)));
+        }
+        if let Some(c) = &f.file_census
+            && !c.objects.is_empty()
+        {
+            let kept = f.drawing_census.as_ref().map(|d| &d.objects);
+            let list: Vec<String> = c
+                .objects
+                .iter()
+                .map(|(k, n)| match kept.and_then(|m| m.get(k)) {
+                    Some(m) => format!("{k} {n} ({m} kept as unmodelled)"),
+                    None => format!("{k} {n}"),
+                })
+                .collect();
+            s.push_str(&format!("- Objects in the file: {}\n", cell(&list.join(", "))));
+        }
+        if !f.read.is_empty() {
+            s.push_str("\n#### Read (file → drawing)\n\n");
+            diff_table(&f.read, s);
+        }
+        for r in &f.routes {
+            s.push_str(&format!("\n#### {}\n\n", r.route));
+            match (&r.save_error, &r.reopen_error) {
+                (Some(e), _) => s.push_str(&format!("- Save: {}: {}\n", status(&r.save_error), cell(e))),
+                (None, Some(e)) => {
+                    s.push_str(&format!("- Save: ok ({})\n", size(r.saved_bytes as u64)));
+                    s.push_str(&format!("- Reopen: {}: {}\n", status(&r.reopen_error), cell(e)));
+                }
+                (None, None) => {
+                    let t = r.comparison.total();
+                    s.push_str(&format!(
+                        "- Saved {} and reopened in {:.1} s: {} lost, {} added, {} changed\n",
+                        size(r.saved_bytes as u64),
+                        r.seconds,
+                        t.lost,
+                        t.added,
+                        t.changed
+                    ));
+                }
+            }
+            if !r.not_written.is_empty() {
+                s.push_str(&format!("- **Lost on save**, not written (the writer can't write these types): {}\n", per_type(&r.not_written)));
+            }
+            if let Some(p) = &r.saved_path {
+                s.push_str(&format!("- Saved file: `{p}`\n"));
+            }
+            for n in &r.reopen_notes {
+                s.push_str(&format!("- Note on reopening: {}\n", cell(n)));
+            }
+            if !r.comparison.is_empty() {
+                s.push('\n');
+                diff_table(&r.comparison, s);
+            }
+        }
+    }
+
+    /// Write `report.md` and `report.json` into `dir` (created if needed); returns their paths.
+    pub fn write(&self, dir: &Path) -> std::io::Result<(PathBuf, PathBuf)> {
+        std::fs::create_dir_all(dir)?;
+        let (md, json) = (dir.join("report.md"), dir.join("report.json"));
+        std::fs::write(&md, self.to_markdown())?;
+        std::fs::write(&json, self.to_json())?;
+        Ok((md, json))
+    }
+}
+
+/// A comparison's differences as a Markdown table (the first [`MD_ROWS`]).
+fn diff_table(c: &Comparison, s: &mut String) {
+    let tallies: Vec<String> = Category::ALL
+        .iter()
+        .filter_map(|cat| {
+            let t = c.tally(*cat);
+            (t.total() > 0).then(|| format!("{} {}/{}/{}", cat.label(), t.lost, t.added, t.changed))
+        })
+        .collect();
+    s.push_str(&format!("By category (lost/added/changed): {}\n\n", tallies.join(" · ")));
+    s.push_str("| Category | Change | Where | Detail |\n|---|---|---|---|\n");
+    for d in c.differences.iter().take(MD_ROWS) {
+        s.push_str(&format!("| {} | {} | {} | {} |\n", d.category.label(), d.change.label(), cell(&clip(&d.place, 80)), cell(&clip(&d.detail, 400))));
+    }
+    let more = c.differences.len().saturating_sub(MD_ROWS).saturating_add(c.omitted);
+    if more > 0 {
+        s.push_str(&format!("\n… and {more} more differences (see report.json; differences beyond the listing limit are only counted).\n"));
+    }
+}

crates/io/src/roundtrip/tests.rs

@@ -0,0 +1,566 @@
+use std::sync::Arc;
+
+use cadcraft_color::Color;
+use cadcraft_doc::foreign::{ForeignExtra, ForeignObject};
+use cadcraft_doc::*;
+use cadcraft_geom::{PolyVertex, Vec2, Vec3};
+
+use super::*;
+
+fn tag(code: i32, value: &str) -> RawTag {
+    RawTag { code, value: value.into() }
+}
+
+fn line(a: (f64, f64), b: (f64, f64)) -> EntityKind {
+    EntityKind::Line(Line { a: Vec3::new(a.0, a.1, 0.0), b: Vec3::new(b.0, b.1, 0.0) })
+}
+
+/// A small drawing touching every part the comparator looks at.
+fn drawing() -> Drawing {
+    let mut d = Drawing::new_metric();
+    d.layers.push(Layer { name: "Walls".into(), color: Color::Index(1), ..Layer::default() });
+    let walls = Common { layer: "Walls".into(), ..Common::default() };
+    d.add(&Space::Model, walls.clone(), line((0.0, 0.0), (10.0, 0.0))).unwrap();
+    d.add(&Space::Model, walls, line((0.0, 5.0), (10.0, 5.0))).unwrap();
+    d.add(&Space::Model, Common::default(), EntityKind::Circle(Circle { center: Vec3::new(3.0, 4.0, 0.0), radius: 2.5 })).unwrap();
+    let tagged = Common { xdata: vec![tag(1001, "ACME"), tag(1000, "door"), tag(1040, "2.5")], ..Common::default() };
+    d.add(
+        &Space::Model,
+        tagged,
+        EntityKind::LwPolyline(LwPolyline {
+            vertices: vec![PolyVertex::new(Vec2::ZERO), PolyVertex::with_bulge(Vec2::new(4.0, 0.0), 0.5), PolyVertex::new(Vec2::new(4.0, 3.0))],
+            closed: true,
+            const_width: 0.0,
+            elevation: 0.0,
+            plinegen: false,
+        }),
+    )
+    .unwrap();
+    d.add(
+        &Space::Model,
+        Common::default(),
+        EntityKind::Unknown(Unknown { dxf_type: "ACME_WIDGET".into(), tags: vec![tag(100, "AcDbEntity"), tag(10, "1.0"), tag(20, "2.0")] }),
+    )
+    .unwrap();
+    d.add(&Space::Paper("Layout1".into()), Common::default(), line((1.0, 1.0), (2.0, 2.0))).unwrap();
+    let mut b = Block::new("DOOR");
+    let h = d.new_handle();
+    b.entities.push(Entity::new(h, line((0.0, 0.0), (0.0, 1.0))));
+    d.blocks.insert("DOOR".into(), Arc::new(b));
+    let f = Arc::make_mut(&mut d.foreign);
+    f.objects.push(ForeignObject { handle: Handle(0x500), kind: "XRECORD".into(), tags: vec![tag(100, "AcDbXrecord"), tag(1, "kept")] });
+    f.dictionary.push(("ACME_DATA".into(), Handle(0x500)));
+    f.classes.push(vec![tag(1, "ACME_WIDGET"), tag(2, "AcmeWidget"), tag(3, "Acme")]);
+    f.records.insert(("LAYER".into(), "WALLS".into()), ForeignExtra { xdata: vec![tag(1001, "ACME"), tag(1070, "3")], ..ForeignExtra::default() });
+    d
+}
+
+fn cmp(a: &Drawing, b: &Drawing) -> Comparison {
+    compare(a, b, &Options::default())
+}
+
+fn only(c: &Comparison, cat: Category, change: Change) -> Vec<&Difference> {
+    c.differences.iter().filter(|d| d.category == cat && d.change == change).collect()
+}
+
+fn first_handle(d: &Drawing, kind: &str) -> Handle {
+    d.model.iter().find(|e| e.kind.dxf_name() == kind).map(|e| e.handle).unwrap()
+}
+
+#[test]
+fn identical_drawings_have_no_differences() {
+    let d = drawing();
+    let c = cmp(&d, &d.clone());
+    assert!(c.is_empty(), "{:#?}", c.differences);
+    assert_eq!(c.total(), Tally::default());
+}
+
+#[test]
+fn geometry_is_compared_by_handle_within_the_tolerance() {
+    let a = drawing();
+    let h = first_handle(&a, "LINE");
+    let mut b = a.clone();
+    b.modify_entity(h, |e| {
+        if let EntityKind::Line(l) = &mut e.kind {
+            l.b.x += 1e-9;
+        }
+    })
+    .unwrap();
+    assert!(cmp(&a, &b).is_empty(), "a change within the tolerance is no difference");
+    b.modify_entity(h, |e| {
+        if let EntityKind::Line(l) = &mut e.kind {
+            l.b.x = 11.0;
+        }
+    })
+    .unwrap();
+    let c = cmp(&a, &b);
+    let g = only(&c, Category::Geometry, Change::Changed);
+    assert_eq!(g.len(), 1, "{:#?}", c.differences);
+    assert!(g[0].place.contains(&h.hex()) && g[0].place.starts_with("Model LINE"), "{:?}", g[0]);
+    assert!(g[0].detail.contains("b.x: 10") && g[0].detail.contains("11"), "{:?}", g[0]);
+    assert_eq!(c.tally(Category::Geometry).changed, 1);
+}
+
+#[test]
+fn angles_a_whole_turn_apart_are_the_same() {
+    let mut a = drawing();
+    let h = a
+        .add(&Space::Model, Common::default(), EntityKind::Arc(cadcraft_doc::Arc { center: Vec3::ZERO, radius: 2.0, start: -0.5, end: 1.0 }))
+        .unwrap();
+    let mut b = a.clone();
+    b.modify_entity(h, |e| {
+        if let EntityKind::Arc(x) = &mut e.kind {
+            x.start += std::f64::consts::TAU;
+            x.end -= std::f64::consts::TAU;
+        }
+    })
+    .unwrap();
+    assert!(cmp(&a, &b).is_empty(), "{:#?}", cmp(&a, &b).differences);
+    // A radius is no angle.
+    b.modify_entity(h, |e| {
+        if let EntityKind::Arc(x) = &mut e.kind {
+            x.radius += std::f64::consts::TAU;
+        }
+    })
+    .unwrap();
+    assert_eq!(cmp(&a, &b).tally(Category::Geometry).changed, 1);
+}
+
+#[test]
+fn properties_and_type_changes_are_reported() {
+    let a = drawing();
+    let h = first_handle(&a, "CIRCLE");
+    let mut b = a.clone();
+    b.modify_entity(h, |e| e.common.color = Color::Index(3)).unwrap();
+    let c = cmp(&a, &b);
+    assert_eq!(only(&c, Category::Properties, Change::Changed).len(), 1, "{:#?}", c.differences);
+    b.modify_entity(h, |e| e.kind = line((0.0, 0.0), (1.0, 1.0))).unwrap();
+    let c = cmp(&a, &b);
+    let g = only(&c, Category::Geometry, Change::Changed);
+    assert!(g.iter().any(|d| d.detail == "type CIRCLE → LINE"), "{:#?}", c.differences);
+}
+
+#[test]
+fn lost_and_added_entities_are_counted_per_type_and_space() {
+    let a = drawing();
+    let h = first_handle(&a, "CIRCLE");
+    let mut b = a.clone();
+    b.remove_entity(h);
+    let mut paper = b.clone();
+    paper.add(&Space::Paper("Layout1".into()), Common::default(), EntityKind::Point(Point { p: Vec3::ZERO, angle: 0.0 })).unwrap();
+    let c = cmp(&a, &paper);
+    let lost = only(&c, Category::Count, Change::Lost);
+    assert_eq!(lost.len(), 1, "{:#?}", c.differences);
+    assert_eq!(lost[0].place, "Model CIRCLE");
+    assert!(lost[0].detail.contains("1 → 0") && lost[0].detail.contains(&h.hex()), "{:?}", lost[0]);
+    let added = only(&c, Category::Count, Change::Added);
+    assert_eq!(added.len(), 1);
+    assert_eq!(added[0].place, "Layout Layout1 POINT");
+    assert_eq!(c.tally(Category::Count), Tally { lost: 1, added: 1, changed: 0 });
+}
+
+#[test]
+fn renumbered_entities_are_matched_by_content() {
+    let a = drawing();
+    let mut b = a.clone();
+    // Every line gets a new handle, in reverse order, and one of them moves.
+    let lines: Vec<Arc<Entity>> = b.model.iter().filter(|e| e.kind.dxf_name() == "LINE").cloned().collect();
+    for (i, e) in lines.iter().rev().enumerate() {
+        b.model.remove(e.handle);
+        let mut e = Entity::clone(e);
+        e.handle = Handle(0x9000 + i as u64);
+        b.model.push(e);
+    }
+    let c = cmp(&a, &b);
+    assert_eq!(c.tally(Category::Count), Tally::default(), "{:#?}", c.differences);
+    assert_eq!(c.tally(Category::Geometry), Tally::default(), "pairs by content, not by order: {:#?}", c.differences);
+    assert_eq!(c.tally(Category::Handle).changed, 2);
+}
+
+#[test]
+fn entities_moving_between_spaces_are_reported() {
+    let a = drawing();
+    let h = first_handle(&a, "CIRCLE");
+    let mut b = a.clone();
+    let e = b.remove_entity(h).unwrap();
+    if let Some(l) = b.layouts.first_mut() {
+        l.entities.push(Entity::clone(&e));
+    }
+    let c = cmp(&a, &b);
+    let moved = only(&c, Category::Handle, Change::Changed);
+    assert_eq!(moved.len(), 1, "{:#?}", c.differences);
+    assert_eq!(moved[0].detail, format!("1 entities moved from Model to Layout Layout1 ({})", h.hex()));
+    assert_eq!(c.tally(Category::Count), Tally::default());
+}
+
+#[test]
+fn unreferenced_anonymous_blocks_dropped_on_save_are_one_row() {
+    let mut a = drawing();
+    for name in ["*D5", "*D6"] {
+        let mut blk = Block::new(name);
+        for _ in 0..2 {
+            let h = a.new_handle();
+            blk.entities.push(Entity::new(h, line((0.0, 0.0), (1.0, 0.0))));
+        }
+        a.blocks.insert(name.into(), Arc::new(blk));
+    }
+    let mut b = a.clone();
+    b.blocks.remove("*D5");
+    b.blocks.remove("*D6");
+    let c = cmp(&a, &b);
+    let lost = only(&c, Category::Count, Change::Lost);
+    assert_eq!(lost.len(), 1, "{:#?}", c.differences);
+    assert_eq!(lost[0].detail, "4 entities of 2 anonymous blocks nothing refers to (e.g. *D5) were not saved");
+    assert_eq!(c.tally(Category::Count).lost, 4);
+    assert_eq!(c.tally(Category::Block), Tally { lost: 2, added: 0, changed: 0 }, "{:#?}", c.differences);
+    assert_eq!(only(&c, Category::Block, Change::Lost).len(), 1);
+}
+
+#[test]
+fn renamed_anonymous_blocks_are_matched_by_their_entities() {
+    let mut a = drawing();
+    let mut blk = Block::new("*D1");
+    let h = a.new_handle();
+    blk.entities.push(Entity::new(h, line((0.0, 0.0), (5.0, 0.0))));
+    a.blocks.insert("*D1".into(), Arc::new(blk.clone()));
+    let key = ("BLOCK_RECORD".to_string(), "*D1".to_string());
+    Arc::make_mut(&mut a.foreign).records.insert(key, ForeignExtra { xdata: vec![tag(1001, "ACME"), tag(1070, "1")], ..ForeignExtra::default() });
+    let mut b = a.clone();
+    b.blocks.remove("*D1");
+    blk.name = "*D7".into();
+    b.blocks.insert("*D7".into(), Arc::new(blk));
+    let f = Arc::make_mut(&mut b.foreign);
+    let extra = f.records.remove(&("BLOCK_RECORD".to_string(), "*D1".to_string())).unwrap();
+    f.records.insert(("BLOCK_RECORD".into(), "*D7".into()), extra);
+    let c = cmp(&a, &b);
+    assert_eq!(c.tally(Category::Count), Tally::default(), "{:#?}", c.differences);
+    assert_eq!(c.tally(Category::Xdata), Tally::default(), "{:#?}", c.differences);
+    assert_eq!(c.tally(Category::Block), Tally { lost: 0, added: 0, changed: 1 }, "{:#?}", c.differences);
+    assert!(c.differences.iter().any(|d| d.detail == "1 anonymous blocks saved under new names (e.g. *D1 → *D7)"), "{:#?}", c.differences);
+}
+
+#[test]
+fn xdata_extension_dictionaries_and_opaque_entities_are_compared() {
+    let a = drawing();
+    let pl = first_handle(&a, "LWPOLYLINE");
+    let unknown = first_handle(&a, "ACME_WIDGET");
+    let mut b = a.clone();
+    b.modify_entity(pl, |e| e.common.xdata.clear()).unwrap();
+    b.modify_entity(unknown, |e| {
+        if let EntityKind::Unknown(u) = &mut e.kind {
+            u.tags.pop();
+        }
+    })
+    .unwrap();
+    Arc::make_mut(&mut b.foreign).records.clear();
+    let mut a2 = a.clone();
+    Arc::make_mut(&mut a2.foreign)
+        .entities
+        .insert(pl, ForeignExtra { xdict: Some(Handle(0x600)), reactors: vec![Handle(0x601)], ..ForeignExtra::default() });
+    let c = cmp(&a2, &b);
+    let x = only(&c, Category::Xdata, Change::Lost);
+    assert_eq!(x.len(), 2, "the entity's and the layer's: {:#?}", c.differences);
+    assert!(x.iter().any(|d| d.place == "LAYER WALLS" && d.detail.starts_with("application ACME")), "{x:#?}");
+    assert!(x.iter().any(|d| d.place.contains("LWPOLYLINE")), "{x:#?}");
+    assert_eq!(c.tally(Category::ExtDict).lost, 2, "the dictionary and a reactor: {:#?}", c.differences);
+    let o = only(&c, Category::Opaque, Change::Changed);
+    assert_eq!(o.len(), 1, "{:#?}", c.differences);
+    assert!(o[0].detail.contains("3 groups → 2"), "{:?}", o[0]);
+    // Numbers in kept groups compare within the tolerance.
+    let mut b2 = a.clone();
+    b2.modify_entity(unknown, |e| {
+        if let EntityKind::Unknown(u) = &mut e.kind {
+            u.tags[1].value = "1.0000000000001".into();
+        }
+    })
+    .unwrap();
+    assert!(cmp(&a, &b2).is_empty());
+}
+
+#[test]
+fn tables_blocks_and_layouts_are_compared_by_name() {
+    let a = drawing();
+    let mut b = a.clone();
+    b.layers.retain(|l| l.name != "Walls");
+    b.linetypes.push(Linetype { name: "Dashed".into(), description: String::new(), pattern: Vec::new() });
+    if let Some(s) = b.text_styles.first_mut() {
+        s.height = 7.0;
+    }
+    b.blocks.remove("DOOR");
+    b.layouts.retain(|l| l.name != "Layout2");
+    let c = cmp(&a, &b);
+    let t = |cat, ch, place: &str| c.differences.iter().any(|d| d.category == cat && d.change == ch && d.place == place);
+    assert!(t(Category::Table, Change::Lost, "LAYER Walls"), "{:#?}", c.differences);
+    assert!(t(Category::Table, Change::Added, "LTYPE Dashed"));
+    assert!(t(Category::Table, Change::Changed, "STYLE Standard"));
+    assert!(t(Category::Block, Change::Lost, "BLOCK DOOR"));
+    assert!(t(Category::Layout, Change::Lost, "LAYOUT Layout2"));
+    // The block's entity is lost with it, counted where it was.
+    assert!(t(Category::Count, Change::Lost, "Block DOOR LINE"), "{:#?}", c.differences);
+    // Names match without case.
+    let mut upper = a.clone();
+    if let Some(l) = upper.layers.iter_mut().find(|l| l.name == "Walls") {
+        l.name = "WALLS".into();
+    }
+    let c = cmp(&a, &upper);
+    assert_eq!(c.tally(Category::Table).lost, 0, "{:#?}", c.differences);
+}
+
+#[test]
+fn unmodelled_objects_dictionary_entries_and_classes_are_compared() {
+    let a = drawing();
+    let mut b = a.clone();
+    let f = Arc::make_mut(&mut b.foreign);
+    f.objects.clear();
+    f.dictionary.clear();
+    f.classes.clear();
+    let c = cmp(&a, &b);
+    assert_eq!(c.tally(Category::Object).lost, 1, "{:#?}", c.differences);
+    assert_eq!(c.tally(Category::Dictionary).lost, 1);
+    assert_eq!(c.tally(Category::Class).lost, 1);
+    // A renumbered object is matched by its groups.
+    let mut b = a.clone();
+    Arc::make_mut(&mut b.foreign).objects[0].handle = Handle(0x700);
+    let c = cmp(&a, &b);
+    assert_eq!(c.tally(Category::Object), Tally::default(), "{:#?}", c.differences);
+    assert_eq!(c.tally(Category::Handle).changed, 1);
+}
+
+#[test]
+fn header_variables_are_compared_without_volatile_ones() {
+    let a = drawing();
+    let mut b = a.clone();
+    b.header.set_f64("LTSCALE", 2.0);
+    b.header.vars.remove("TEXTSIZE");
+    b.header.set_str("HANDSEED", "FFFF");
+    b.header.set_str("TDUPDATE", "2461000.5");
+    // 1 and 1.0 are the same value.
+    b.header.set_i64("DIMSCALE", 1);
+    let c = cmp(&a, &b);
+    assert!(only(&c, Category::Header, Change::Changed).iter().any(|d| d.place == "$LTSCALE"), "{:#?}", c.differences);
+    assert!(only(&c, Category::Header, Change::Lost).iter().any(|d| d.place == "$TEXTSIZE"));
+    assert_eq!(c.tally(Category::Header).total(), 2, "{:#?}", c.differences);
+}
+
+#[test]
+fn the_listing_limit_keeps_counting() {
+    let a = drawing();
+    let mut b = a.clone();
+    for i in 0..20 {
+        b.header.set_i64(&format!("USERI{i}"), i);
+    }
+    let c = compare(&a, &b, &Options { max_differences: 5, ..Options::default() });
+    assert_eq!(c.differences.len(), 5);
+    assert_eq!(c.omitted, 15);
+    assert_eq!(c.tally(Category::Header).added, 20);
+}
+
+#[test]
+fn census_counts_what_a_dxf_holds() {
+    let dxf = "  0\nSECTION\n  2\nTABLES\n  0\nTABLE\n  2\nLAYER\n  0\nLAYER\n  2\n0\n  0\nLAYER\n  2\nA\n  0\nENDTAB\n  0\nENDSEC\n\
+               0\nSECTION\n  2\nBLOCKS\n  0\nBLOCK\n  2\n*Model_Space\n  0\nENDBLK\n  0\nBLOCK\n  2\n*Paper_Space\n  0\nLINE\n  0\nENDBLK\n\
+               0\nBLOCK\n  2\nB1\n  0\nCIRCLE\n  0\nPOLYLINE\n  0\nVERTEX\n  0\nSEQEND\n  0\nENDBLK\n  0\nENDSEC\n\
+               0\nSECTION\n  2\nENTITIES\n  0\nLINE\n  0\nINSERT\n  0\nATTRIB\n  0\nSEQEND\n  0\nENDSEC\n\
+               0\nSECTION\n  2\nOBJECTS\n  0\nDICTIONARY\n  0\nXRECORD\n  0\nXRECORD\n  0\nENDSEC\n  0\nEOF\n";
+    let c = census(dxf.as_bytes()).unwrap();
+    assert_eq!(c.entities.get("LINE"), Some(&2));
+    assert_eq!(c.entities.get("INSERT"), Some(&1));
+    assert!(!c.entities.contains_key("ATTRIB") && !c.entities.contains_key("SEQEND"));
+    assert_eq!(c.block_entities.get("CIRCLE"), Some(&1));
+    assert_eq!(c.block_entities.get("POLYLINE"), Some(&1));
+    assert_eq!(c.blocks, 1);
+    assert_eq!(c.tables.get("LAYER"), Some(&2));
+    assert_eq!(c.objects.get("XRECORD"), Some(&2));
+    let _ = census(b"");
+}
+
+#[test]
+fn read_census_reports_entities_the_reader_left_out() {
+    let d = drawing();
+    let mut file = census_of(&d);
+    *file.entities.entry("LINE".into()).or_default() += 2;
+    *file.entities.entry("WEIRD".into()).or_default() += 1;
+    let c = compare_census(&file, &d, &Options::default());
+    let lost = only(&c, Category::Read, Change::Lost);
+    assert_eq!(c.tally(Category::Read).lost, 3, "{:#?}", c.differences);
+    assert!(lost.iter().any(|x| x.place == "spaces LINE" && x.detail == "LINE: 5 in the file, 3 read"), "{lost:#?}");
+    assert!(compare_census(&census_of(&d), &d, &Options::default()).is_empty());
+}
+
+/// The round trips of a synthetic drawing through CADCraft's own DXF and DWG writers: whatever
+/// they lose shows in the comparison. Plain geometry, layers, xdata and unmodelled objects must
+/// come back from DXF.
+#[cfg(not(target_arch = "wasm32"))]
+#[test]
+fn dxf_round_trip_of_plain_geometry_is_lossless() {
+    use corpus::{Format, round_trip};
+    let mut d = drawing();
+    // What the comparator should see in a round trip: the drawing as read from its own DXF.
+    d = crate::read(crate::write(&d, "a.dxf").unwrap().as_slice(), "a.dxf").unwrap();
+    let r = round_trip(&d, Format::Dxf, Format::Dxf, &Options::default(), None);
+    assert!(r.ok(), "{r:#?}");
+    let core = [Category::Count, Category::Geometry, Category::Table, Category::Xdata, Category::Object, Category::Opaque];
+    for cat in core {
+        assert_eq!(r.comparison.tally(cat), Tally::default(), "{cat:?}: {:#?}", r.comparison.differences);
+    }
+    let dwg = round_trip(&d, Format::Dxf, Format::Dwg, &Options::default(), None);
+    assert!(dwg.save_error.is_none() && dwg.reopen_error.is_none(), "{dwg:#?}");
+    assert!(dwg.saved_bytes > 0);
+    // Through DWG, only the custom-class entity is lost (gap FF1: acadrust drops other programs'
+    // entities without DWG data of their own) — and the comparator says so.
+    let lost: Vec<&str> = dwg.comparison.differences.iter().filter(|d| d.category == Category::Count).map(|d| d.place.as_str()).collect();
+    assert_eq!(lost, ["Model ACME_WIDGET"], "{:#?}", dwg.comparison.differences);
+    assert_eq!(dwg.comparison.tally(Category::Geometry), Tally::default(), "{:#?}", dwg.comparison.differences);
+}
+
+#[cfg(not(target_arch = "wasm32"))]
+#[test]
+fn hostile_files_are_reported_not_crashed() {
+    use corpus::{CorpusOptions, check_bytes};
+    let opts = CorpusOptions::default();
+    for (bytes, name) in [
+        (&b""[..], "empty.dxf"),
+        (b"AC1032 not really a dwg file at all", "fake.dwg"),
+        (b"  0\nSECTION\n  2\nENTITIES\n  0\nLINE\n 10\nnan\n 20\n1e400\n  0\nENDSEC\n  0\nEOF\n", "odd.dxf"),
+        (&[0xFFu8; 64][..], "noise.dxf"),
+    ] {
+        let f = check_bytes(bytes, name, &opts);
+        assert_eq!(f.path, name);
+        if f.opened() {
+            assert!(!f.routes.is_empty());
+        } else {
+            assert!(f.routes.is_empty() && f.open_error.as_deref().is_some_and(|e| !e.is_empty()), "{f:#?}");
+        }
+    }
+}
+
+#[cfg(not(target_arch = "wasm32"))]
+#[test]
+fn report_renders_as_markdown_and_json() {
+    use corpus::{CorpusOptions, CorpusReport, check_bytes};
+    let d = drawing();
+    let dxf = crate::write(&d, "a.dxf").unwrap();
+    let opts = CorpusOptions::default();
+    let report = CorpusReport {
+        corpus: "corpus".into(),
+        generated: "2026-10-21 09:30 UTC".into(),
+        cadcraft_version: "0".into(),
+        abs_tol: 1e-6,
+        rel_tol: 1e-9,
+        files: vec![check_bytes(&dxf, "synthetic/a.dxf", &opts), check_bytes(b"AC1032 broken", "dwg/broken.dwg", &opts)],
+    };
+    let md = report.to_markdown();
+    assert!(md.starts_with("# CADCraft round-trip report"), "{md}");
+    assert!(md.contains("| `synthetic/a.dxf` | DXF AC1032 |"), "{md}");
+    assert!(md.contains("dxf → dxf") && md.contains("dxf → dwg"), "{md}");
+    assert!(md.contains("| `dwg/broken.dwg` | DWG AC1032 |") && md.contains("**failed**"), "{md}");
+    assert!(md.contains("## By category"));
+    let json: serde_json::Value = serde_json::from_str(&report.to_json()).unwrap();
+    assert_eq!(json["files"][0]["path"], "synthetic/a.dxf");
+    assert_eq!(json["files"][0]["routes"][0]["route"], "dxf → dxf");
+    assert!(json["files"][1]["openError"].as_str().is_some_and(|e| !e.is_empty()), "{json}");
+}
+
+#[cfg(not(target_arch = "wasm32"))]
+#[test]
+fn a_folder_it_cannot_read_is_reported_not_skipped() {
+    let w = corpus::walk(std::path::Path::new("no/such/corpus/folder"), &[]);
+    assert!(w.files.is_empty());
+    assert_eq!(w.unreadable.len(), 1, "{:?}", w.unreadable);
+}
+
+#[test]
+fn corpus_files_walks_folders_and_skips_the_report() {
+    use corpus::corpus_files;
+    let dir = std::env::temp_dir().join(format!("cadcraft-corpus-walk-{}", std::process::id()));
+    let _ = std::fs::remove_dir_all(&dir);
+    for p in ["dxf/a.DXF", "dwg/b.dwg", "dwg/sub/c.dwg", "notes.txt", "report/a.roundtrip.dxf", ".git/x.dxf"] {
+        let path = dir.join(p);
+        std::fs::create_dir_all(path.parent().unwrap()).unwrap();
+        std::fs::write(&path, b"x").unwrap();
+    }
+    let files = corpus_files(&dir, &[dir.join("report")]);
+    let rel: Vec<String> = files.iter().map(|f| corpus::relative(f, &dir)).collect();
+    assert_eq!(rel, ["dwg/b.dwg", "dwg/sub/c.dwg", "dxf/a.DXF"]);
+    let _ = std::fs::remove_dir_all(&dir);
+}
+
+/// The real-file corpus: set `CADCRAFT_CORPUS_DIR` to a folder of DXF/DWG files (a checkout of
+/// `storytold/cadcraft-corpus`) to open, save and reopen each one and write `report.md` and
+/// `report.json` to `CADCRAFT_CORPUS_REPORT` (default: the corpus folder's `report`). Skipped
+/// when the variable is unset, so CI passes without the corpus. It reports, it doesn't fail on
+/// differences; set `CADCRAFT_CORPUS_STRICT=1` to fail when a file doesn't open, save or reopen,
+/// or when a writer leaves records out of a saved file ([`corpus::CorpusReport::strict_failures`]).
+#[cfg(not(target_arch = "wasm32"))]
+#[test]
+fn corpus_round_trip_report() {
+    let Some(dir) = std::env::var_os("CADCRAFT_CORPUS_DIR").map(std::path::PathBuf::from) else {
+        eprintln!("CADCRAFT_CORPUS_DIR is not set: corpus round trip skipped");
+        return;
+    };
+    let out = std::env::var_os("CADCRAFT_CORPUS_REPORT").map_or_else(|| dir.join("report"), std::path::PathBuf::from);
+    let files = corpus::corpus_files(&dir, std::slice::from_ref(&out));
+    let report = corpus::run(&dir, &files, &corpus::CorpusOptions::default(), &mut |i, n, rel| eprintln!("[{}/{n}] {rel}", i + 1));
+    let (md, json) = report.write(&out).unwrap();
+    eprintln!("corpus report: {} and {}", md.display(), json.display());
+    if std::env::var("CADCRAFT_CORPUS_STRICT").is_ok_and(|v| v == "1") {
+        let failed = report.strict_failures();
+        assert!(failed.is_empty(), "corpus failures:\n{}", failed.join("\n"));
+    }
+}
+
+/// A type the DWG writer can't write is reported as lost on save, per type: in the route's
+/// `not_written`, on the lost entity rows, in the Markdown and as a strict failure. A type it
+/// writes (OLE2FRAME, kept by CADCraft without modelling it) is none of these.
+#[cfg(not(target_arch = "wasm32"))]
+#[test]
+fn types_the_dwg_writer_cannot_write_are_reported_lost() {
+    use corpus::{CorpusReport, FileReport, Format, round_trip};
+    let unknown = |t: &str, tags: Vec<RawTag>| EntityKind::Unknown(Unknown { dxf_type: t.into(), tags });
+    let mut d = Drawing::new_metric();
+    let widget = vec![tag(100, "AcDbEntity"), tag(100, "AcmeWidget"), tag(10, "1.0"), tag(20, "2.0"), tag(30, "0.0")];
+    d.add(&Space::Model, Common::default(), unknown("ACME_WIDGET", widget.clone())).unwrap();
+    d.add(&Space::Model, Common::default(), unknown("ACME_WIDGET", widget)).unwrap();
+    // Its data has no DWG wrapper, which is where a DWG keeps the corners: the reader's defaults.
+    let ole = ["100", "AcDbEntity", "8", "0", "100", "AcDbOle2Frame", "70", "2", "10", "1.0", "20", "1.0", "30", "0.0", "11", "0.0", "21", "0.0"]
+        .chunks(2)
+        .map(|g| tag(g[0].parse().unwrap(), g[1]))
+        .chain(
+            ["31", "0.0", "71", "2", "72", "0", "73", "0", "90", "4", "310", "C0FFEE00", "1", "OLE"]
+                .chunks(2)
+                .map(|g| tag(g[0].parse().unwrap(), g[1])),
+        )
+        .collect();
+    d.add(&Space::Model, Common::default(), unknown("OLE2FRAME", ole)).unwrap();
+    let d = crate::read(&crate::write(&d, "a.dxf").unwrap(), "a.dxf").unwrap();
+
+    let r = round_trip(&d, Format::Dxf, Format::Dwg, &Options::default(), None);
+    assert!(r.ok(), "{r:#?}");
+    assert_eq!(r.not_written, [("ACME_WIDGET".to_string(), 2)].into_iter().collect());
+    let lost: Vec<&Difference> = r.comparison.differences.iter().filter(|x| x.category == Category::Count && x.change == Change::Lost).collect();
+    assert_eq!(lost.len(), 1, "{lost:#?}");
+    assert!(lost[0].place.ends_with("ACME_WIDGET") && lost[0].detail.contains("not written: the DWG writer can't write ACME_WIDGET"), "{lost:#?}");
+    // The OLE2FRAME is written (only the CLASS record AutoCAD's files have is added).
+    assert!(
+        !r.comparison.differences.iter().any(|x| x.category != Category::Class && x.place.contains("OLE2FRAME")),
+        "{:#?}",
+        r.comparison.differences
+    );
+    assert!(!r.lossless());
+
+    let report =
+        CorpusReport { files: vec![FileReport { path: "a.dxf".into(), routes: vec![r], ..FileReport::default() }], ..CorpusReport::default() };
+    assert_eq!(report.strict_failures(), ["a.dxf (dxf → dwg): lost on save, not written: 2 ACME_WIDGET"]);
+    let md = report.to_markdown();
+    assert!(md.contains("- **Lost on save** (records a writer can't write): 2 ACME_WIDGET (dxf → dwg)"), "{md}");
+    assert!(md.contains("| dxf → dwg | ok | ok | **2** |"), "{md}");
+    assert!(md.contains("- **Lost on save**, not written (the writer can't write these types): 2 ACME_WIDGET"), "{md}");
+
+    // DXF keeps everything: nothing lost on save, nothing to fail on.
+    let r = round_trip(&d, Format::Dxf, Format::Dxf, &Options::default(), None);
+    assert!(r.not_written.is_empty());
+    let report =
+        CorpusReport { files: vec![FileReport { path: "a.dxf".into(), routes: vec![r], ..FileReport::default() }], ..CorpusReport::default() };
+    assert!(report.strict_failures().is_empty());
+    assert!(report.to_markdown().contains("- Lost on save (records a writer can't write): none"));
+}

xtask/src/main.rs

@@ -23,8 +23,10 @@ commands:
   test            cargo nextest run (cargo test without nextest) plus the doctests
   i18n            print the strings each interface language shows in English (never fails)
   ci              fmt --check, clippy -D warnings, test, i18n, assets, layers, wasm (stops at first failure)
-  corpus [--download]
-                  show where test corpora live; --download fetches PngSuite into corpus/pngsuite
+  corpus [DIR] [--out DIR] [--keep-saved] [--same-format]
+                  open -> save -> reopen every DXF/DWG of a real-file corpus (DIR, default
+                  $CADCRAFT_CORPUS_DIR or ../cadcraft-corpus) and write report.md + report.json
+                  (default DIR/report); runs `cadcraft-cli roundtrip` in release mode
   stats [--exact] count tests and lines per crate (--exact: ask the test harness via `-- --list`)
   parity          recompute the AutoCAD menu-breadth checklist in docs/parity-checklist.md
   ico <out.ico> <in.png>...
@@ -41,7 +43,7 @@ fn main() -> ExitCode {
         Some("test") => cmd_test(),
         Some("i18n") => cmd_i18n(),
         Some("ci") => cmd_ci(),
-        Some("corpus") => cmd_corpus(rest.contains(&"--download")),
+        Some("corpus") => cmd_corpus(&rest),
         Some("stats") => stats::run(&root(), rest.contains(&"--exact")),
         Some("parity") => parity::run(&root()),
         Some("ico") => ico::run(&rest),
@@ -226,7 +228,47 @@ fn cmd_ci() -> Result<(), String> {
     Ok(())
 }
 
-fn cmd_corpus(_download: bool) -> Result<(), String> {
-    println!("Test corpora live in storytold/cadcraft-corpus (planned); never commit large binaries here.");
-    Ok(())
+/// `corpus [DIR] [options]`: the round-trip report over a real-file corpus (gaps FF1, FF5). The
+/// corpus lives outside this repo (storytold/cadcraft-corpus); its files are never committed here.
+fn cmd_corpus(args: &[&str]) -> Result<(), String> {
+    // `cadcraft-cli` runs from the repository root: relative paths are made absolute against the
+    // folder `cargo xtask` was run from first, so they mean what the user typed.
+    let here = std::env::current_dir().map_err(|e| format!("current folder: {e}"))?;
+    let mut pass: Vec<String> = Vec::new();
+    let mut dir: Option<String> = None;
+    let mut it = args.iter();
+    while let Some(a) = it.next() {
+        match *a {
+            "--out" => {
+                let v = it.next().ok_or_else(|| format!("{a} needs a value"))?;
+                pass.extend([a.to_string(), here.join(v).display().to_string()]);
+            }
+            "--tolerance" => {
+                let v = it.next().ok_or_else(|| format!("{a} needs a value"))?;
+                pass.extend([a.to_string(), v.to_string()]);
+            }
+            s if s.starts_with("--out=") => {
+                let v = s.trim_start_matches("--out=");
+                pass.extend(["--out".to_string(), here.join(v).display().to_string()]);
+            }
+            s if s.starts_with("--") => pass.push(s.to_string()),
+            s if dir.is_none() => dir = Some(s.to_string()),
+            _ => return Err("corpus takes one DIR".into()),
+        }
+    }
+    let dir = dir
+        .or_else(|| std::env::var("CADCRAFT_CORPUS_DIR").ok())
+        .unwrap_or_else(|| root().parent().map(|p| p.join("cadcraft-corpus").display().to_string()).unwrap_or_else(|| "../cadcraft-corpus".into()));
+    let dir = here.join(&dir);
+    if !dir.is_dir() {
+        return Err(format!(
+            "no corpus at `{}`: clone storytold/cadcraft-corpus next to this repo, set CADCRAFT_CORPUS_DIR or pass DIR \
+             (never commit corpus files to this repo)",
+            dir.display()
+        ));
+    }
+    let dir = dir.display().to_string();
+    let mut c = cargo();
+    c.args(["run", "--release", "-q", "-p", "cadcraft-cli", "--", "roundtrip", &dir]).args(&pass);
+    run(c, &format!("cadcraft-cli roundtrip {dir} {}", pass.join(" ")))
 }