Fix rate limiter to throttle every call, not just after N

Confirmed via AI Studio dashboard: gemini-3.6-flash free tier caps at 5 RPM. Old limiter allowed unthrottled bursts up to 15 calls before pausing, which 429'd. New limiter enforces a minimum delay between every single call.

app.py

@@ -24,7 +24,7 @@ DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
 @spaces.GPU
 def _zerogpu_startup_check():
     # This Space runs on ZeroGPU hardware but the agent below only makes
-    # network calls (Claude API, web search) and never touches CUDA.
+    # network calls (Gemini API, web search) and never touches CUDA.
     # ZeroGPU requires at least one @spaces.GPU function to be declared,
     # so this no-op satisfies that check without spending any GPU quota
     # (it is never actually invoked).
@@ -41,22 +41,24 @@ If you are asked for a comma separated list, apply the above rules to each eleme
 
 class RateLimiter:
     """
-    Counts agent LLM calls (one per step) and sleeps for a fixed duration
-    once a threshold is reached, then resets the counter. Used as a
-    step_callback to stay under the model provider's rate limits across a
-    full run of ~20 GAIA questions, each potentially taking several steps.
+    Enforces a minimum delay between successive agent LLM calls (one call per
+    step) as a step_callback, to stay under the model provider's
+    requests-per-minute limit. Gemini's free tier for gemini-3.6-flash caps
+    at 5 RPM, so the default here (12s + margin) targets that; override via
+    RATE_LIMIT_SECONDS_BETWEEN_CALLS if your tier/model allows more.
     """
-    def __init__(self, calls_per_wait: int = 15, seconds_to_wait: int = 30):
-        self.calls_per_wait = calls_per_wait
-        self.seconds_to_wait = seconds_to_wait
-        self._call_count = 0
+    def __init__(self, min_seconds_between_calls: float = 13.0):
+        self.min_seconds_between_calls = min_seconds_between_calls
+        self._last_call_at: float | None = None
 
     def __call__(self, memory_step: ActionStep, agent: CodeAgent) -> None:
-        self._call_count += 1
-        if self._call_count >= self.calls_per_wait:
-            print(f"Rate limiter: {self.calls_per_wait} calls reached, sleeping {self.seconds_to_wait}s")
-            time.sleep(self.seconds_to_wait)
-            self._call_count = 0
+        now = time.monotonic()
+        if self._last_call_at is not None:
+            wait = self.min_seconds_between_calls - (now - self._last_call_at)
+            if wait > 0:
+                print(f"Rate limiter: sleeping {wait:.1f}s to stay under the RPM limit")
+                time.sleep(wait)
+        self._last_call_at = time.monotonic()
 
 
 # --- Basic Agent Definition ---
@@ -78,7 +80,7 @@ class BasicAgent:
                 "collections", "statistics", "datetime", "io", "openpyxl", "PIL",
             ],
             max_steps=12,
-            step_callbacks=[RateLimiter()],
+            step_callbacks=[RateLimiter(float(os.getenv("RATE_LIMIT_SECONDS_BETWEEN_CALLS", "13.0")))],
         )
         print("BasicAgent initialized.")