feat(metrics): C3 — Prometheus counters for L0/L1/format-loss/retry

services/quality/pipeline.py

@@ -48,6 +48,13 @@ def run_l0_check(
         issues={"internal_error": 1},
     )
 
+    # Best-effort import of metrics; never let it break the check.
+    try:
+        from middleware.metrics import record_l0_result, record_l0_error
+    except Exception:
+        record_l0_result = None
+        record_l0_error = None
+
     try:
         result = evaluate_document(source_chunks, translated_chunks, target_lang)
     except Exception as e:
@@ -60,6 +67,11 @@ def run_l0_check(
             error_type=type(e).__name__,
             elapsed_ms=elapsed_ms,
         )
+        if record_l0_error:
+            try:
+                record_l0_error(file_type=file_extension or "unknown")
+            except Exception:
+                pass
         return empty
 
     elapsed_ms = round((time.time() - start) * 1000, 2)
@@ -75,6 +87,14 @@ def run_l0_check(
         issues=result.issues,
         elapsed_ms=elapsed_ms,
     )
+    if record_l0_result:
+        try:
+            record_l0_result(
+                passed=bool(result.passed),
+                file_type=file_extension or "unknown",
+            )
+        except Exception:
+            pass
     return result
 
 
@@ -145,6 +165,7 @@ async def run_l1_check(
             reason="insufficient_chunks_or_all_flagged",
             chunk_count=len(source_chunks),
         )
+        _record_l1_metric(verdict="skip", model="none")
         return skip
 
     # Get the judge
@@ -157,6 +178,7 @@ async def run_l1_check(
             job_id=job_id,
             reason="no_judge_configured",
         )
+        _record_l1_metric(verdict="skip", model="none")
         return skip
 
     # Get the language name for the prompt
@@ -173,6 +195,7 @@ async def run_l1_check(
             error=str(e)[:200],
             error_type=type(e).__name__,
         )
+        _record_l1_metric(verdict="error", model="unknown")
         return L1Result(
             verdict="skip",
             chunks_evaluated=0, chunks_passed=0, chunks_failed=0,
@@ -195,9 +218,46 @@ async def run_l1_check(
         cost_estimate_usd=result.cost_estimate_usd,
         log_only=log_only,
     )
+
+    # Record Prometheus metric (best-effort, never breaks the job)
+    duration_s = None
+    if result.elapsed_ms is not None:
+        try:
+            duration_s = float(result.elapsed_ms) / 1000.0
+        except Exception:
+            duration_s = None
+    _record_l1_metric(
+        verdict=result.verdict or "skip",
+        model=result.model_used or "unknown",
+        duration_seconds=duration_s,
+        cost_usd=result.cost_estimate_usd,
+    )
     return result
 
 
+def _record_l1_metric(
+    verdict: str,
+    model: str = "unknown",
+    duration_seconds: float = None,
+    cost_usd: float = None,
+) -> None:
+    """Best-effort Prometheus metric emission for L1.
+
+    Never raises — if the metrics module fails to import or the
+    counter is unavailable, we silently skip.
+    """
+    try:
+        from middleware.metrics import record_l1_verdict
+        record_l1_verdict(
+            verdict=verdict,
+            model=model,
+            duration_seconds=duration_seconds,
+            cost_usd=cost_usd,
+        )
+    except Exception:
+        pass
+
+
 def make_judge_from_env_safe() -> Optional[LLMJudge]:
     """Read env vars and build a judge, or return None if not configured.