Fix unreliable capture of standard/error outputs

Fix potentially incomplete output capture when running unit tests, as
seen with e.g. GNU Guile.  Two changes:

- Drain standard output and error output pipes
- Remove hard-coded stdout/stderr truncation to 2kB

Author: rudolf-adamkovic

evaluation/src/containerized_eval.py

@@ -78,9 +78,6 @@ def eval_string_script(language, program):
         f.write(program.encode("utf-8"))
         f.flush()
         result = eval_script(Path(f.name))
-        # Only save the first 2K of output from the running program. Any futher
-        # output is very likely an exceptionally long stack trace or a long
-        # series of prints.
         if type(result["stdout"]) == bytes:
             result["stdout"] = result["stdout"].decode("utf-8", errors="ignore")
         if result["stdout"] is None:
@@ -93,8 +90,8 @@ def eval_string_script(language, program):
         assert type(result["stderr"]) == str
         return {
             "program": program,
-            "stdout": result['stdout'].replace("!!int", "")[:2048],
-            "stderr": result['stderr'][:2048],
+            "stdout": result['stdout'].replace("!!int", ""),
+            "stderr": result['stderr'],
             "exit_code": result['exit_code'],
             "status": result['status']
         }

evaluation/src/safe_subprocess/__init__.py

@@ -47,7 +47,9 @@ def run(
         stdout=subprocess.PIPE,
         stderr=subprocess.PIPE,
         start_new_session=True,
-        bufsize=MAX_BYTES_PER_READ,
+        # Use raw IO because BufferedReader returns None for partial
+        # reads on non-blocking pipes which can lead to data loss.
+        bufsize=0,
         cwd=cwd
     )
     set_nonblocking(p.stdout)
@@ -61,6 +63,7 @@ def run(
     stderr_saved_bytes = []
     stdout_bytes_read = 0
     stderr_bytes_read = 0
+    exit_code = None
 
     for _ in range(max_iterations):
         this_stdout_read = p.stdout.read(MAX_BYTES_PER_READ)
@@ -74,9 +77,21 @@ def run(
             stderr_saved_bytes.append(this_stderr_read)
             stderr_bytes_read += len(this_stderr_read)
         exit_code = p.poll()
-        if exit_code is not None:
-            break
-        time.sleep(SLEEP_BETWEEN_READS)
+        if exit_code is None:
+            time.sleep(SLEEP_BETWEEN_READS)
+        else:
+            # After exit, keep looping without sleeping to drain any output
+            # remaining in the pipe buffers; stop once both pipes are empty.
+            this_stdout_read = p.stdout.read(MAX_BYTES_PER_READ)
+            this_stderr_read = p.stderr.read(MAX_BYTES_PER_READ)
+            if this_stdout_read is not None and stdout_bytes_read < max_output_size:
+                stdout_saved_bytes.append(this_stdout_read)
+                stdout_bytes_read += len(this_stdout_read)
+            if this_stderr_read is not None and stderr_bytes_read < max_output_size:
+                stderr_saved_bytes.append(this_stderr_read)
+                stderr_bytes_read += len(this_stderr_read)
+            if not this_stdout_read and not this_stderr_read:
+                break
 
     try:
         # Kills the process group. Without this line, test_fork_once fails.