Fix remaining review issues: null safety, compressed count, batch tokens

- M8: Fix null prompt_tokens_details in streaming.py (2 more instances)
- Issue 2: requests_compressed now counts by tokens_saved > 0 only,
not savings_percent > 0 (which could be negative from pipeline overhead)
- H3: Initialize original_tokens=0 before try block in Google batch handler
so error handler has a valid value
- L4: Google batch handler now looks up model context limit via
openai_provider.get_context_limit() instead of hardcoding 128000

headroom/proxy/cost.py

@@ -240,7 +240,7 @@ def build_session_summary(
             if entry.model and "count_tokens" in entry.model:
                 uncompressed_reasons["passthrough"] += 1
                 continue
-            if entry.tokens_saved > 0 and entry.savings_percent > 0:
+            if entry.tokens_saved > 0:
                 compressed_requests.append(
                     {
                         "savings_pct": round(entry.savings_percent, 1),

headroom/proxy/handlers/batch.py

@@ -134,9 +134,15 @@ class BatchHandlerMixin:
                 continue
 
             # Apply optimization
+            original_tokens = 0  # Set before try so error handler can use it
+            optimized_tokens = 0
             try:
-                # Default context limit for most models
-                context_limit = 128000
+                # Look up model context limit, fall back to 128K
+                context_limit = (
+                    self.openai_provider.get_context_limit(model)
+                    if hasattr(self, "openai_provider")
+                    else 128000
+                )
 
                 # Use OpenAI pipeline (similar message format after conversion)
                 result = self.openai_pipeline.apply(
@@ -219,8 +225,7 @@ class BatchHandlerMixin:
                 )
                 # Pass through unchanged on failure — count original as optimized
                 compressed_requests.append(batch_req)
-                # original_tokens may be unbound if pipeline failed before producing result
-                # Just skip the token accounting for this failed request
+                total_optimized_tokens += original_tokens  # 0 if pipeline never ran
 
         # Update body with compressed requests
         body["batch"]["input_config"]["requests"]["requests"] = compressed_requests

headroom/proxy/handlers/streaming.py

@@ -80,7 +80,7 @@ class StreamingMixin:
                         usage["input_tokens"] = chunk_usage.get("prompt_tokens", 0)
                         usage["output_tokens"] = chunk_usage.get("completion_tokens", 0)
                         # OpenAI has cached tokens in prompt_tokens_details
-                        details = chunk_usage.get("prompt_tokens_details", {})
+                        details = chunk_usage.get("prompt_tokens_details") or {}
                         usage["cache_read_input_tokens"] = details.get("cached_tokens", 0)
 
                 elif provider == "gemini":
@@ -164,7 +164,7 @@ class StreamingMixin:
                     if chunk_usage:
                         usage_found["input_tokens"] = chunk_usage.get("prompt_tokens", 0)
                         usage_found["output_tokens"] = chunk_usage.get("completion_tokens", 0)
-                        details = chunk_usage.get("prompt_tokens_details", {})
+                        details = chunk_usage.get("prompt_tokens_details") or {}
                         usage_found["cache_read_input_tokens"] = details.get("cached_tokens", 0)
 
                 elif provider == "gemini":