From dc9a29410190185a8759ca56cc9609ae46aeea5b Mon Sep 17 00:00:00 2001
From: Cyd <cyd@falloutshelterarcade.com>
Date: Mon, 6 Jul 2026 13:41:16 -0500
Subject: [PATCH] Writer thread survives write faults instead of dying mid-run

A single failed port write killed the paced writer thread permanently
while the port still reported open, so every later byte (ACKs, the
CheckRequest handshake) queued forever: wire tap showed 2 of 3 bytes of
a stray ButtonPressed at t=54ms, then 1167 unanswered CheckRequests.
The trigger was com0com flow control with the game side not yet open --
the third byte blocked past WriteTimeout, threw, and the catch returned.

A real UART cannot wedge; it shifts bits into the line whether or not
anyone listens. On a write fault, drop the stalled byte plus the stale
backlog, log the stall/recovery transition once, and keep the writer
alive -- TX resumes as soon as the host drains its end.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 src/VRio.Core/Device/VRioSerialService.cs | 34 +++++++++++++++++++++--
 1 file changed, 31 insertions(+), 3 deletions(-)

diff --git a/src/VRio.Core/Device/VRioSerialService.cs b/src/VRio.Core/Device/VRioSerialService.cs
index 4d33cfe..16fb36b 100644
--- a/src/VRio.Core/Device/VRioSerialService.cs
+++ b/src/VRio.Core/Device/VRioSerialService.cs
@@ -18,6 +18,12 @@ namespace VRio.Core.Device;
 /// monotonic slot deadline (<c>slot = max(prevSlot + period, now)</c>), so
 /// the stream averages the true baud rate without bursting after idle.</para>
 ///
+/// <para>A write fault never kills the writer — a real UART streams into an
+/// unterminated line rather than blocking. If the virtual wire's far side
+/// stops draining (peer end closed, write timeout), the stalled byte and the
+/// queued backlog are dropped and transmission resumes with the next fresh
+/// packet once the host reads again.</para>
+///
 /// <para>RIOJoy pulses DTR for 50 ms when it opens its end (the board-reset
 /// handshake); through a null modem that arrives here as a DSR blip, which is
 /// surfaced via <see cref="HostHandshake"/> so the UI can show that a host
@@ -201,6 +207,7 @@ public sealed class VRioSerialService : IDisposable
     {
         var one = new byte[1];
         long slot = Stopwatch.GetTimestamp();
+        bool txHealthy = true; // log stall/recovery transitions, not every byte
 
         while (_running)
         {
@@ -226,12 +233,33 @@ public sealed class VRioSerialService : IDisposable
             try
             {
                 port.Write(one, 0, 1);
+                if (!txHealthy)
+                {
+                    txHealthy = true;
+                    Logged?.Invoke("TX recovered — host is draining the wire again");
+                }
             }
             catch (Exception ex) when (ex is IOException or InvalidOperationException or TimeoutException)
             {
-                if (_running)
-                    Logged?.Invoke($"Write failed: {ex.Message}");
-                return;
+                if (!_running)
+                    return;
+                // A real UART cannot wedge: it shifts bits onto the line whether
+                // or not anyone is listening. A failed write means the virtual
+                // wire's far side stopped draining (peer end closed), so the
+                // queued backlog is already stale — drop it and keep serving
+                // fresh traffic; writes land again once the host reads its end.
+                int dropped;
+                lock (_txGate)
+                {
+                    dropped = _txQueue.Count;
+                    _txQueue.Clear();
+                }
+                if (txHealthy)
+                {
+                    txHealthy = false;
+                    Logged?.Invoke($"TX stalled ({ex.Message.TrimEnd('.')}) — dropped {dropped + 1} stale byte(s), writer stays alive");
+                }
+                continue;
             }
 
             // If the wait overshot its slot, pace the next byte from the