Fix research pipeline budget tracking

src/nsct/api/rest_research.py

@@ -66,7 +66,7 @@ DEPTH_CONFIGS: dict[str, DepthConfig] = {
         max_pages_per_domain=2,
         max_total_download_bytes=500_000,
         max_llm_requests=15,
-        max_research_duration_seconds=300,
+        max_research_duration_seconds=1_800,
         max_context_per_llm_call=16_000,
     ),
     "normal": DepthConfig(
@@ -76,7 +76,7 @@ DEPTH_CONFIGS: dict[str, DepthConfig] = {
         max_pages_per_domain=5,
         max_total_download_bytes=2_000_000,
         max_llm_requests=30,
-        max_research_duration_seconds=300,
+        max_research_duration_seconds=3_600,
         max_context_per_llm_call=24_000,
     ),
     "deep": DepthConfig(
@@ -86,7 +86,7 @@ DEPTH_CONFIGS: dict[str, DepthConfig] = {
         max_pages_per_domain=10,
         max_total_download_bytes=5_000_000,
         max_llm_requests=60,
-        max_research_duration_seconds=600,
+        max_research_duration_seconds=7_200,
         max_context_per_llm_call=32_000,
     ),
 }

src/nsct/crawler/manager.py

@@ -167,7 +167,10 @@ class CrawlerManager:
     @staticmethod
     def _error_doc(url: str, error: str) -> NormalizedDocument:
         """Create a NormalizedDocument representing an error."""
-        doc = NormalizedDocument(
+        # Keep the error result structurally identical to a successfully
+        # normalized document.  ``from_text`` hashes the (empty) content, so
+        # downstream consumers can safely rely on content_hash being present.
+        doc = NormalizedDocument.from_text(
             url=url,
             text="",
             title="",
@@ -176,7 +179,6 @@ class CrawlerManager:
                 "error": error,
                 "content_type": "error",
             },
-            content_hash="",
             extraction_tool="",
         )
         return doc

src/nsct/orchestration/budget.py

@@ -74,6 +74,7 @@ class BudgetTracker:
     def __init__(self, config: HardBudgetConfig) -> None:
         self._config = config
         self._start_time = time.monotonic()
+        self._last_time_recorded = self._start_time
         # internal counters
         self._counters: dict[str, int | float] = {
             "search_queries": 0,
@@ -132,12 +133,11 @@ class BudgetTracker:
     def record_time_elapsed(self) -> None:
         """Tick the internal elapsed-time clock.
 
-        Adds the time since the tracker was created (or the last call)
-        to the elapsed duration.
+        Adds only the time since the previous call to the elapsed duration.
         """
-        self.update_research_duration(
-            time.monotonic() - self._start_time
-        )
+        now = time.monotonic()
+        self.update_research_duration(now - self._last_time_recorded)
+        self._last_time_recorded = now
 
     def increment_context_tokens(self, n: int) -> None:
         """Add *n* tokens to the current LLM context window counter."""

src/nsct/orchestration/orchestrator.py

@@ -345,8 +345,8 @@ class ResearchOrchestrator:
         ResearchRun
             Das erstellte Run-Objekt.
         """
-        # Budget init
-        self._budget_tracker.increment_llm_requests(1)  # planner call
+        # The planner call itself is accounted for in ``_step_planning``.
+        # Starting a run performs no LLM request.
         self._budget_tracker.increment_search(1)  # initial search plan
 
         # ResearchRun erstellen
@@ -724,7 +724,26 @@ class ResearchOrchestrator:
                 self._claims = []
                 return {"success": True, "data": {"claims": []}, "claim_count": 0}
 
-            self._budget_tracker.increment_llm_requests(len(self._sources))
+            # Error documents are retained in ``_sources`` for transparent
+            # reporting, but they must never consume a claim-extraction LLM
+            # request.  Reserve one request for the synthesis step as well,
+            # otherwise a quick run can use its entire LLM budget here and
+            # fail deterministically before synthesis.
+            extractable_sources = [source for source in self._sources if not source.get("error")]
+            usage = self._budget_tracker.get_usage()
+            remaining_llm_requests = self._budget_config.max_llm_requests - int(
+                usage["max_llm_requests"]["usage"]
+            )
+            synthesis_reserve = 1
+            max_sources_for_extraction = max(0, remaining_llm_requests - synthesis_reserve)
+            extractable_sources = extractable_sources[:max_sources_for_extraction]
+
+            if not extractable_sources:
+                logger.warning("No successful sources or LLM budget for claim extraction")
+                self._claims = []
+                return {"success": True, "data": {"claims": []}, "claim_count": 0}
+
+            self._budget_tracker.increment_llm_requests(len(extractable_sources))
 
             # Stage 5: Claim Extraction
             try:
@@ -734,7 +753,7 @@ class ResearchOrchestrator:
                     llm_provider=llm_provider,
                     config=self._config,
                     research_run_id=self._run.id if self._run else uuid4(),
-                    sources=self._sources,
+                    sources=extractable_sources,
                 )
                 self._claims = await extractor.extract()
             except NameError: