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

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

@@ -0,0 +1,29 @@
+# Correctness Lens
+
+Attack the changed behavior, not every subsystem in the repository.
+
+- Trace new error paths to the caller. Inject the ignored/wildcard variants and
+  verify they cannot become success, not-found, or a plausible default.
+- Exercise zero/empty/missing, maximum, and exact-boundary inputs for every
+  changed count, size, index, page limit, or optional value.
+- For aggregation/quorum changes, test exactly quorum and quorum-minus-one with
+  mixed disk errors and nil/placeholder entries.
+- For listing/pagination, test `n == max`, `n == max + 1`, delimiter folding,
+  continuation markers, and object/prefix name collisions.
+- For EC/read/streaming changes, inject failure after partial output and verify
+  the client receives an error rather than a clean truncated body. Assert exact
+  bytes and length.
+- For multipart/object commits, fail before/after rename and cleanup; committed
+  data must remain readable and pre-commit cleanup must not destroy parts.
+- For version/index ordering, test `len - 1`, `len`, equal timestamps, missing
+  versions, and deterministic tie-breaking.
+- For directory-object behavior, trace `__XLDIR__` at the store layer; branches
+  below the layer that sees trailing slashes are dead.
+- For binary UUID metadata, absent, empty, and nil all mean no value. Never send
+  nil/empty `versionId` to an unversioned tier.
+- For agent rules/skill routers, test a trigger matrix covering ordinary
+  inquiry, low-risk implementation, explicit review, high-risk code, PR
+  creation, release, and post-PR monitoring. Each case must select only the
+  intended workflow and retain required safety/authorization boundaries.
+
+Null verdicts name only the probes relevant to the diff.