in-flight shows admitted/held: bare 96 with gate 2 read as broken

The counter was taken just past the GLOBAL semaphore (lifted to 96 in
auto mode so the gate is the sole limiter) but BEFORE the pool gate --
so 94 gate-queued workers counted as in-flight. The gate now pushes
its actually-admitted count and the bar shows 'admitted/held'
(e.g. in-flight 2/96 = 2 really hitting the server, 94 queued on the
gate). Verified: admitted never exceeds the gate limit.

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

evalharness/model/pool.py

@@ -203,6 +203,18 @@ class AdaptiveGate:
             except Exception:
                 pass
 
+    def push_inflight(self) -> None:
+        """Tell the bar how many requests the gate has ACTUALLY admitted
+        (bar shows 'admitted/held' -- a bare held count with gate 2 read
+        as 'the gate is broken')."""
+        rep = (self.adapter.extra or {}).get('progress_reporter')
+        fn = getattr(rep, 'set_admitted', None)
+        if fn is not None:
+            try:
+                fn(self._inflight)
+            except Exception:
+                pass
+
     # ---- gate semantics ----
     async def acquire(self) -> None:
         loop = asyncio.get_running_loop()
@@ -230,9 +242,11 @@ class AdaptiveGate:
             finally:
                 self._cond.release()
         self._inflight += 1
+        self.push_inflight()
 
     def release(self, ok: bool) -> None:
         self._inflight = max(0, self._inflight - 1)
+        self.push_inflight()
         if ok:
             self._interval_ok += 1
             self._level_ok += 1

evalharness/progress/rich_terminal.py

@@ -60,6 +60,7 @@ class RichTerminalProgress:
         self.started = 0.0
         self.current_started = 0.0
         self.inflight = 0
+        self.admitted = None  # requests past the pool gate (None = no gate)
         self.restored = 0     # checkpoint head start (drives the '+N new' marker)
         self.heartbeat_task = None
 
@@ -195,7 +196,7 @@ class RichTerminalProgress:
             return
         self.inflight += 1
         self.current_started = time.monotonic()
-        self.progress.update(self.task_id, inflight=self.inflight, cur='0s')
+        self.progress.update(self.task_id, inflight=self._inflight_txt(), cur='0s')
 
     def set_retries(self, n: int):
         """Show the retry count on the bar (from the adapter's attempt)."""
@@ -207,6 +208,21 @@ class RichTerminalProgress:
         if self.task_id is not None:
             self.progress.update(self.task_id, gate=f'gate {n}')
 
+    def _inflight_txt(self) -> str:
+        # 'held' counts workers past the global semaphore; when a pool gate
+        # is active most of them are QUEUED on it -- 'admitted' is what
+        # actually hits the server. Showing a bare 96 with gate 2 read as
+        # 'the gate is not working'
+        if self.admitted is None:
+            return str(self.inflight)
+        return f'{self.admitted}/{self.inflight}'
+
+    def set_admitted(self, n: int):
+        """Pooled runs: requests actually admitted by the gate."""
+        self.admitted = n
+        if self.task_id is not None:
+            self.progress.update(self.task_id, inflight=self._inflight_txt())
+
     def rollback(self):
         if self.disabled:
             return
@@ -214,7 +230,7 @@ class RichTerminalProgress:
         just decrement the in-flight count, no success/fail bookkeeping."""
         self.inflight = max(0, self.inflight - 1)
         if self.task_id is not None:
-            self.progress.update(self.task_id, inflight=self.inflight, cur='0s')
+            self.progress.update(self.task_id, inflight=self._inflight_txt(), cur='0s')
 
     def advance(self, success: bool = True):
         if self.disabled:
@@ -233,7 +249,7 @@ class RichTerminalProgress:
             advance=1,
             new=self._new_txt(completed),
             rate=f"{fresh / elapsed:.2f}",
-            inflight=self.inflight, cur='0s',
+            inflight=self._inflight_txt(), cur='0s',
             elapsed=_fmt(elapsed),
             eta=_fmt((task.total - completed) * elapsed / fresh)
                  if fresh and task.total and task.total > completed else '-',