diff --git a/Cargo.lock b/Cargo.lock
index 549cde179..1a7e8860b 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -1270,6 +1270,7 @@ dependencies = [
  "cucumber",
  "e2e-report",
  "futures",
+ "libc",
  "portable-pty",
  "rand 0.9.4",
  "reqwest 0.13.4",
diff --git a/tests/e2e-cucumber/Cargo.toml b/tests/e2e-cucumber/Cargo.toml
index 36aecc620..268fb36bd 100644
--- a/tests/e2e-cucumber/Cargo.toml
+++ b/tests/e2e-cucumber/Cargo.toml
@@ -26,7 +26,11 @@ path = "src/bin/fake-tailscale.rs"
 axum.workspace = true
 cucumber = { version = "0.23", features = ["output-json", "output-junit"] }
 e2e-report = { path = "../../crates/e2e-report" }
-# Builds the paced download fixture's chunked response body (`stream::unfold`).
+# Two users: it builds the paced download fixture's chunked response body
+# (`stream::unfold`), and it provides `FutureExt::catch_unwind`, which catches a
+# panic escaping the cucumber run (see `src/harness_diagnostics.rs`) — plain
+# `std::panic::catch_unwind` cannot span an `.await`. cucumber already pulls this
+# in, so it adds no crate to the tree.
 futures = "0.3"
 # Drive the interactive dash TUI black-box: spawn the real `rocm` binary under a
 # pseudo-terminal (`portable-pty`, cross-platform openpty/ConPTY) and parse the
@@ -62,6 +66,12 @@ vt100 = "0.16"
 # dev edges by design — so keeping it here needs no allowlist entry.
 rocm-core = { path = "../../crates/rocm-core" }
 
+# `fcntl(F_GETFL)` only, to read (never modify) the `O_NONBLOCK` state of the
+# suite's own stdio for the diagnostics log (see `src/harness_diagnostics.rs`).
+# Already a workspace dependency, so this adds no new crate to the tree.
+[target.'cfg(unix)'.dependencies]
+libc.workspace = true
+
 [[test]]
 name = "e2e"
 harness = false
diff --git a/tests/e2e-cucumber/src/harness_diagnostics.rs b/tests/e2e-cucumber/src/harness_diagnostics.rs
new file mode 100644
index 000000000..f7cbe4492
--- /dev/null
+++ b/tests/e2e-cucumber/src/harness_diagnostics.rs
@@ -0,0 +1,368 @@
+// Copyright © Advanced Micro Devices, Inc., or its affiliates.
+//
+// SPDX-License-Identifier: MIT
+
+//! Make the harness's own death legible instead of silent.
+//!
+//! The self-hosted WSL2 lane fails by stopping mid-suite with nothing to go on: the
+//! last line is a passing step, then `cargo`'s `error: test failed` with no
+//! `Caused by:` — which narrows it to exit code 101, an unwinding panic — and no
+//! panic message anywhere in the job log. `report.json` and `junit.xml` are left at
+//! zero bytes and `report.html` is never written, so the artifact says nothing
+//! either. Every guess at the cause costs a full CI round trip on a scarce runner.
+//!
+//! **Why the message is missing.** cucumber's runner replaces the panic hook with
+//! an empty one for the whole run and restores it afterwards
+//! (`cucumber-0.23.0/src/runner/basic.rs:929`, restored at `:1070`) — deliberately,
+//! so a step's panic is reported by the writer rather than printed twice. A panic
+//! *raised by the writer itself* escapes that path: it unwinds out of `run()` past
+//! the restore, with the silencing hook still installed. Nothing prints, and the
+//! exit code is all that is left. That is not specific to any lane; it is why every
+//! failure of this shape has been unreadable.
+//!
+//! So this does not install a hook — one would be taken away again on the next
+//! line. [`run_or_record`] catches the panic where it escapes, at the `run()`
+//! boundary, and records it two ways: to stderr, which the silenced hook would have
+//! written, and to a log inside the results directory the lane already uploads,
+//! which survives even if stderr is the thing that broke. The panic is then
+//! re-raised, so the process still dies with 101 and nothing downstream has to
+//! learn a new signal.
+//!
+//! The same log records the state of the standard streams at start-up and again at
+//! the panic, so a descriptor that changed underneath the process is visible rather
+//! than inferred. This only observes the streams; it does not modify them.
+//!
+//! **What this does not recover.** Only the panic *message* survives: the
+//! `file:line` and any backtrace are handed to the panic hook, not carried in the
+//! unwind payload, and the hook installed at that moment is cucumber's empty one.
+//! So a one-line entry with no location is everything there is, not a truncation.
+//! And only a panic that unwinds out of `run()` reaches this boundary — one inside
+//! a fixture server's task or thread (`http_server::spawn` drops its
+//! `JoinHandle`) is contained there, and while the empty hook is installed it
+//! prints nothing either. The step that needed that server then fails on a
+//! symptom — a refused connection, a timeout — with no trace of the cause.
+//!
+//! The whole boundary lives in the library target rather than inline in the
+//! `harness = false` cucumber test binary, so its logic gets real `#[test]`
+//! coverage — the custom harness never executes plain `#[test]` functions placed
+//! inside it. Same reason as [`crate::panic_capture`].
+
+use std::future::Future;
+use std::io::Write as _;
+use std::path::Path;
+use std::pin::Pin;
+
+use futures::FutureExt as _;
+
+/// Name of the log inside the results directory. Picked up by the lane's existing
+/// `upload-artifact` of the whole directory, so nothing in CI needs to change.
+pub const LOG_NAME: &str = "harness-diagnostics.log";
+
+/// Describe the standard streams: their `O_NONBLOCK` state and what they point at.
+///
+/// Read-only. Best-effort and infallible — a description that cannot be obtained is
+/// reported as such, because this runs on a path where failing would destroy the
+/// very diagnostic it exists to produce.
+#[must_use]
+pub fn describe_std_streams() -> Vec<String> {
+    #[cfg(unix)]
+    {
+        use std::os::fd::{AsFd as _, AsRawFd as _, BorrowedFd};
+
+        #[allow(unsafe_code)] // libc FFI
+        fn flags(fd: BorrowedFd<'_>) -> String {
+            // SAFETY: `fd` is a valid borrowed descriptor; `F_GETFL` takes no
+            // pointer arguments.
+            let raw = unsafe { libc::fcntl(fd.as_raw_fd(), libc::F_GETFL) };
+            if raw < 0 {
+                return format!("flags unreadable ({})", std::io::Error::last_os_error());
+            }
+            let blocking = if raw & libc::O_NONBLOCK == 0 {
+                "blocking"
+            } else {
+                "NON-BLOCKING"
+            };
+            format!("{blocking} (F_GETFL=0o{raw:o})")
+        }
+
+        // `/proc/self/fd/N` names the open file the descriptor currently refers to
+        // ("pipe:[12345]", a tty, a path). Two snapshots naming different targets
+        // is the signature of a descriptor being replaced underneath the process.
+        // Linux-only despite the `cfg(unix)` above: elsewhere there is no `/proc`,
+        // so this half reads `target unknown (..)` and only the flags are reported.
+        fn target(n: i32) -> String {
+            std::fs::read_link(format!("/proc/self/fd/{n}")).map_or_else(
+                |e| format!("target unknown ({e})"),
+                |p| p.display().to_string(),
+            )
+        }
+
+        let stdin = std::io::stdin();
+        let stdout = std::io::stdout();
+        let stderr = std::io::stderr();
+        vec![
+            format!("stdin  -> {} [{}]", target(0), flags(stdin.as_fd())),
+            format!("stdout -> {} [{}]", target(1), flags(stdout.as_fd())),
+            format!("stderr -> {} [{}]", target(2), flags(stderr.as_fd())),
+        ]
+    }
+    #[cfg(not(unix))]
+    {
+        vec!["stream description is POSIX-only; not collected on this platform".to_owned()]
+    }
+}
+
+/// Append a section to the diagnostics log, creating it if needed.
+///
+/// Silently gives up on any I/O error: a diagnostic must never be the reason a run
+/// fails, and there is nowhere left to report the failure of the reporting channel.
+fn append(dir: &Path, heading: &str, lines: &[String]) {
+    let Ok(mut file) = std::fs::OpenOptions::new()
+        .create(true)
+        .append(true)
+        .open(dir.join(LOG_NAME))
+    else {
+        return;
+    };
+    let _ = writeln!(file, "=== {heading} ===");
+    for line in lines {
+        let _ = writeln!(file, "{line}");
+    }
+    let _ = writeln!(file);
+    let _ = file.flush();
+}
+
+/// Record the run's starting conditions.
+///
+/// Written unconditionally — not only when something looks wrong. A successful run
+/// leaves the baseline the next failing run is read against, and "checked, nothing
+/// to report" is a different fact from "never ran".
+pub fn record_startup(dir: &Path) {
+    append(dir, "harness start", &describe_std_streams());
+}
+
+/// Record a panic that escaped the cucumber run, to the log and to stderr.
+///
+/// Both destinations on purpose: stderr is what a reader of the job log expects and
+/// is what the silenced hook would have produced, while the log survives a stderr
+/// that is not reaching anyone. `writeln!` rather than `eprintln!` so a failing
+/// stderr yields a missing line rather than a second panic on the way out.
+pub fn record_fatal_panic(dir: &Path, message: &str) {
+    record_fatal_panic_to(dir, message, &mut std::io::stderr());
+}
+
+/// [`record_fatal_panic`] with its second destination injected, so the text written
+/// there can be asserted on.
+///
+/// Split out purely for that: with `std::io::stderr()` hard-coded, deleting the
+/// write left every test in this module green, which is the one destination the
+/// lane this exists for actually reads. The pointer to the log is the part a reader
+/// needs and the part most likely to rot, so it is what the test pins.
+fn record_fatal_panic_to(dir: &Path, message: &str, out: &mut dyn std::io::Write) {
+    let mut lines = vec![format!("message: {message}")];
+    lines.extend(describe_std_streams());
+    append(dir, "fatal panic (escaped the cucumber run)", &lines);
+
+    let _ = writeln!(
+        out,
+        "E2E suite aborted by a panic inside the cucumber run: {message}\n\
+         (cucumber silences the panic hook while running, so this would otherwise \
+         print nothing; see {LOG_NAME} in the results artifact)"
+    );
+}
+
+/// Drive `fut` to completion, recording any panic that escapes it before re-raising.
+///
+/// This is the boundary described in the module docs: the last point at which the
+/// payload of a writer panic still exists, because cucumber's silencing hook is
+/// still installed as it unwinds past the restore.
+///
+/// The panic is **re-raised, never absorbed** — `resume_unwind` diverges, so the
+/// process still dies with exit 101 exactly as it did before. A diagnostic that
+/// swallowed a failure would turn a red run green, which is strictly worse than the
+/// silence it replaces.
+///
+/// `std::panic::catch_unwind` cannot span an `.await`, hence `FutureExt`.
+///
+/// Returns a boxed future rather than being an `async fn`: the cucumber run it
+/// wraps is enormous, and an `async fn` would store it inline and hand the caller a
+/// future of the same size (`clippy::large_futures`). Allocating it once here keeps
+/// that detail out of every call site.
+pub fn run_or_record<'a, F>(dir: &'a Path, fut: F) -> Pin<Box<dyn Future<Output = F::Output> + 'a>>
+where
+    F: Future + 'a,
+{
+    Box::pin(async move {
+        let guarded = Box::pin(std::panic::AssertUnwindSafe(fut).catch_unwind());
+        match guarded.await {
+            Ok(output) => output,
+            Err(payload) => {
+                record_fatal_panic(dir, &crate::panic_capture::panic_message(&payload));
+                std::panic::resume_unwind(payload);
+            }
+        }
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// The stream snapshot the module promises at *both* ends of the log.
+    ///
+    /// Asserted per-section rather than over the whole file: `record_startup` also
+    /// emits one, so a whole-log `contains` would still pass if the panic-time
+    /// snapshot were dropped — which is the headline claim (comparing the two is
+    /// how a descriptor replaced mid-run becomes visible).
+    fn assert_has_stream_snapshot(section: &str, which: &str, log: &str) {
+        #[cfg(unix)]
+        let (needle, what) = ("stdout ->", "a stream line");
+        #[cfg(not(unix))]
+        let (needle, what) = ("POSIX-only", "the POSIX-only placeholder");
+        assert!(
+            section.contains(needle),
+            "the {which} section carries no stream snapshot ({what} expected):\n{log}"
+        );
+    }
+
+    /// Start-up state and a later fatal panic have to land in the same file, in
+    /// order, so one read of the artifact shows what changed between them.
+    #[test]
+    fn the_startup_baseline_and_a_later_panic_share_one_log() {
+        let dir = tempfile::tempdir().expect("failed to create a temp dir");
+        record_startup(dir.path());
+        record_fatal_panic(dir.path(), "failed to write into terminal: boom");
+
+        let log = std::fs::read_to_string(dir.path().join(LOG_NAME)).expect("no log written");
+        let start = log.find("=== harness start ===").expect("no start section");
+        let panic_at = log.find("=== fatal panic").expect("no panic section");
+        assert!(start < panic_at, "sections out of order:\n{log}");
+        assert!(
+            log.contains("message: failed to write into terminal: boom"),
+            "{log}"
+        );
+
+        let (startup_section, panic_section) = log.split_at(panic_at);
+        assert_has_stream_snapshot(startup_section, "start-up", &log);
+        assert_has_stream_snapshot(panic_section, "panic", &log);
+    }
+
+    /// Every stream is named even when nothing is wrong, so a later reader can tell
+    /// "checked, was fine" from "never collected" — and where the flags are POSIX
+    /// only, the log has to say that rather than go quiet, which reads the same as
+    /// a collection that never ran.
+    #[test]
+    fn the_stream_description_accounts_for_every_standard_stream() {
+        let described = describe_std_streams().join("\n");
+        assert!(!described.is_empty(), "the description must never be empty");
+
+        #[cfg(unix)]
+        for stream in ["stdin", "stdout", "stderr"] {
+            assert!(
+                described.contains(stream),
+                "{stream} missing from: {described}"
+            );
+        }
+        #[cfg(not(unix))]
+        assert!(
+            described.contains("POSIX-only"),
+            "a platform without these flags must say so, got: {described}"
+        );
+    }
+
+    /// A destination that cannot be written to must not take the run down with it —
+    /// this runs on the path that is already handling a fatal panic.
+    #[test]
+    fn an_unwritable_destination_is_survived_silently() {
+        let dir = tempfile::tempdir().expect("failed to create a temp dir");
+        let missing = dir.path().join("does").join("not").join("exist");
+        record_startup(&missing);
+        record_fatal_panic(&missing, "boom");
+        assert!(!missing.exists(), "the helper must not create the tree");
+
+        // Prove those calls were real work that the unwritable path swallowed, and
+        // not a no-op: without this, an `append` that wrote nothing anywhere would
+        // satisfy the assertion above just as well.
+        record_startup(dir.path());
+        assert!(
+            dir.path().join(LOG_NAME).exists(),
+            "a writable destination must still be written"
+        );
+    }
+
+    /// The boundary's whole purpose: the message survives, and the failure still
+    /// fails. Absorbing the panic would turn a red run green.
+    #[tokio::test]
+    async fn a_panic_escaping_the_run_is_recorded_and_re_raised() {
+        let dir = tempfile::tempdir().expect("failed to create a temp dir");
+
+        // The future yields a value on its non-panicking path, as the real one
+        // does. Without that, the block's type is `!`, which makes "swallow the
+        // panic and return something" a compile error rather than a test failure —
+        // and an assertion that no reachable mutation can trip is not a guard.
+        let escaped = std::panic::AssertUnwindSafe(run_or_record(dir.path(), async {
+            assert!(
+                !std::hint::black_box(true),
+                "failed to write into terminal: injected"
+            );
+            "a summary that is never produced"
+        }))
+        .catch_unwind()
+        .await;
+
+        assert!(
+            escaped.is_err(),
+            "the panic must propagate, not be absorbed into a normal return"
+        );
+        let log = std::fs::read_to_string(dir.path().join(LOG_NAME)).expect("no log written");
+        assert!(
+            log.contains("message: failed to write into terminal: injected"),
+            "the escaping panic's message must reach the log:\n{log}"
+        );
+    }
+
+    /// The second destination is the one the lane actually reads.
+    ///
+    /// stderr is what reaches the job log, and the suite lane this boundary exists
+    /// for publishes no artifact at all — so the log is unreachable there and this
+    /// line is the whole diagnostic. Left unasserted, deleting the write kept every
+    /// other test in this module green, which would quietly restore exactly the
+    /// silence the change removes.
+    ///
+    /// Pins the pointer to the log as well as the message: a reader who gets this
+    /// line still needs to be told where the fuller record is, and that is the half
+    /// most likely to rot.
+    #[test]
+    fn the_panic_is_announced_on_the_second_destination_too() {
+        let dir = tempfile::tempdir().expect("failed to create a temp dir");
+        let mut out = Vec::new();
+
+        record_fatal_panic_to(dir.path(), "failed to write into terminal: boom", &mut out);
+
+        let written = String::from_utf8(out).expect("the announcement must be UTF-8");
+        assert!(
+            written.contains("failed to write into terminal: boom"),
+            "the panic message must reach the second destination, got: {written}"
+        );
+        assert!(
+            written.contains(LOG_NAME),
+            "the announcement must point at the fuller record, got: {written}"
+        );
+    }
+
+    /// A run that completes normally passes straight through — the boundary must
+    /// not alter the value or leave a panic section behind on a healthy run.
+    #[tokio::test]
+    async fn a_run_that_does_not_panic_is_passed_through_untouched() {
+        let dir = tempfile::tempdir().expect("failed to create a temp dir");
+
+        let summary = run_or_record(dir.path(), async { "the real summary" }).await;
+
+        assert_eq!(summary, "the real summary");
+        let log = std::fs::read_to_string(dir.path().join(LOG_NAME)).unwrap_or_default();
+        assert!(
+            !log.contains("fatal panic"),
+            "a healthy run must not record a panic:\n{log}"
+        );
+    }
+}
diff --git a/tests/e2e-cucumber/src/lib.rs b/tests/e2e-cucumber/src/lib.rs
index f82bde9f6..97ae7e2e1 100644
--- a/tests/e2e-cucumber/src/lib.rs
+++ b/tests/e2e-cucumber/src/lib.rs
@@ -4,6 +4,7 @@
 
 pub mod capability;
 pub mod expectation;
+pub mod harness_diagnostics;
 pub mod http_server;
 pub mod installer_fixture;
 pub mod loopback_http;
diff --git a/tests/e2e-cucumber/tests/e2e.rs b/tests/e2e-cucumber/tests/e2e.rs
index e68c34b49..082ebdeef 100644
--- a/tests/e2e-cucumber/tests/e2e.rs
+++ b/tests/e2e-cucumber/tests/e2e.rs
@@ -1173,6 +1173,10 @@ async fn main() {
     use e2e_cucumber::monotonic_clock::MonotonicClockWriter;
 
     let dir = results_dir();
+
+    // Baseline for the diagnostics log the lane uploads; see the `run()` boundary
+    // below and `harness_diagnostics` for what it is answering.
+    e2e_cucumber::harness_diagnostics::record_startup(&dir);
     let json_file =
         std::fs::File::create(dir.join("report.json")).expect("failed to create report.json");
     let junit_file =
@@ -1257,7 +1261,11 @@ async fn main() {
         } else {
             64
         };
-    let summary = E2eWorld::cucumber()
+    // Wrapped, not hooked — see `harness_diagnostics::run_or_record`, which this is
+    // handed to below. cucumber replaces the panic hook with an empty one for the
+    // whole run, so a panic raised by the *writer* unwinds out of here with that
+    // silencing hook still installed and prints nothing at all.
+    let cucumber_run = E2eWorld::cucumber()
         .max_concurrent_scenarios(max_concurrent)
         // Record the scenario name on the World before each scenario so every
         // `rocm` invocation can be tied back to its scenario for the coverage
@@ -1342,8 +1350,8 @@ async fn main() {
                 }
                 run
             }
-        })
-        .await;
+        });
+    let summary = e2e_cucumber::harness_diagnostics::run_or_record(&dir, cucumber_run).await;
 
     // Generate the HTML report before exiting so the artifact still uploads on
     // failure.