docs(agents): streamline instruction routing (#6358)

.agents/skills/adversarial-validation/references/performance.md

@@ -0,0 +1,20 @@
+# Performance Lens
+
+- For added clones/allocations on request/object/block paths, quantify copied
+  data and frequency. Recommend borrowing, move, `Bytes`/`Arc`, `Cow`, or
+  capacity reservation only for a concrete repeated cost.
+- Route every new sync/flush through the durability-mode and bucket override
+  gates; mode `none` must not pay the new fsync.
+- Keep blocking filesystem/CPU work off async runtime threads, but do not split
+  one small operation into many `spawn_blocking` round trips.
+- Measure lock hold time across I/O and compare acquisition order for ABBA.
+- Keep cleanup, extra stat/rename, and diagnostics out of the PUT commit critical
+  section when they need not be there.
+- Detect per-item serial I/O/RPC in batch APIs and accidental quadratic scans;
+  use a gate or bounded concurrency when the concrete fan-out warrants it.
+- Count buffer growth and byte copies in EC/bitrot paths; preserve pool gauge
+  balance and avoid repeated metadata decode/fetch per object.
+- Repetitive success logs stay at `trace`; metrics/instrumentation on hot paths
+  require an existing gate.
+- Claims of no impact on PUT/GET/commit/erasure paths need relevant benchmark or
+  A/B evidence, especially for 4 KiB objects.