Idle watchdog back to 300s default; begin_bench clears the stale bar

- watchdog: 10s was the user's experiment; back to 5min default
  (EVALHARNESS_PULL_IDLE_S overrides). Zero-byte detection itself is
  sound: select(fd, timeout) + any-chunk reset, unit-verified.
- a FAILED bench left its 'scoring 0/2' bar on screen through the
  NEXT bench's entire dataset download; begin_bench() relabels to
  '<name> · loading' and zeroes counters at each bench start.

Co-Authored-By: Claude <noreply@anthropic.com>

evalharness/cli.py

@@ -745,6 +745,9 @@ def _cmd_eval_run(args) -> int:
         try:
             if _shared_reporter is not None and total_runs > 1:
                 _shared_reporter.set_bench_tag(f'[{i + 1}/{total_runs}]')
+            if _shared_reporter is not None:
+                _shared_reporter.begin_bench(name)   # clear the previous
+                # bench's stale scoring bar before this one's dataset load
 
             def _emit(msg, _i=i, _n=name):
                 if _shared_reporter is not None:

evalharness/progress/rich_terminal.py

@@ -162,6 +162,26 @@ class RichTerminalProgress:
             self.heartbeat_task.cancel()
         self.heartbeat_task = asyncio.create_task(self._heartbeat())
 
+    def begin_bench(self, name: str):
+        """New benchmark starting: immediately re-label the bar to its
+        loading phase. Without this the PREVIOUS bench's stale scoring
+        state (e.g. a failed '0/2') stayed on screen through the whole
+        dataset download of the next one."""
+        if self.disabled or self.task_id is None:
+            return
+        self.bench_name = name
+        self._last_phase = 'loading'
+        self._scoring_for = None
+        self.started = time.monotonic()
+        self.inflight = 0
+        self.restored = 0
+        self.admitted = None
+        self.progress.update(
+            self.task_id,
+            description=f'[green]{self.bench_tag}{name} · loading[/green]',
+            total=1, completed=0, new='', rate='0.00', inflight=0,
+            cur='0s', elapsed='0s', eta='-', retries=0)
+
     def set_scoring(self, done: int, total: int):
         """Retarget the SAME bar to the scoring phase: generation is finished
         and its filled 100% state is stale -- now the bar refills with judged

evalharness/sandbox/prefetch.py

@@ -166,7 +166,7 @@ def _pull_one(image: str, verbose: bool = True) -> None:
         # override with EVALHARNESS_PULL_IDLE_S for slow-negotiating mirrors.
         import os as _os2
 
-        _IDLE_S = float(_os2.environ.get('EVALHARNESS_PULL_IDLE_S', '10'))
+        _IDLE_S = float(_os2.environ.get('EVALHARNESS_PULL_IDLE_S', '300'))
 
         def _iter_lines():
             nonlocal buf