test(rocke): optimize arch-domain reproduction scheduling

dnn-providers/hip-kernel-provider/rocke/TESTING.md

@@ -421,6 +421,17 @@ skip; a partial or internally inconsistent LLVM installation fails. Flavor
 auto-detection errors, including unavailable COMGR evidence, also remain
 failures during this migration.
 
+The reproduction scheduler preserves exhaustive coverage while avoiding
+redundant compiler startup. It batches declaration discovery, preflights each
+target once, and combines expected-positive calls per architecture. The
+committed artifact schedules those batches but cannot make them pass: every call
+must survive compile, link, and target assembly, and a failed batch is bisected
+to individual verdicts. Expected-negative cells remain individual and use
+explicit operand witnesses. Independent work shares a quota-aware bounded
+worker pool, with `--jobs` available for an explicit site override.
+Deterministic expected crashes are not retried; unexpected or unclassified
+outcomes still receive a contention-free retry.
+
 The fast structure and generator units remain source-only in
 `platform/tests/tools/test_arch_domain_artifact.py`. Normal pytest and the
 `ROCKE_BACKEND=both` pytest pass do not re-run the compiler sweep. The dedicated

dnn-providers/hip-kernel-provider/rocke/platform/tests/README.md

@@ -102,6 +102,21 @@ source tool and committed source column, while the installed entry invokes a
 packaged test-only copy of the tool and the column in the relocated `rocke`
 package. This matches the library-side CI model: CMake stages every input, CTest
 owns execution, and no test reconstructs or reaches back into the source tree.
+
+Reproduction uses the committed column only to schedule work, never as its
+verdict. Declaration recognition is batched through `opt` and recursively split
+on failure. Each compiler/architecture pair receives an intrinsic-free target
+preflight. Calls for expected-positive cells are combined per architecture; a
+failed module is recursively split until ordinary single-cell probes identify
+every changed result. Expected-negative cells remain individual, and every
+positive call still passes through compile, link, and target assembly. Special
+integer and `immarg` operands come from explicit `ProbeSpec` witnesses rather
+than speculative value sweeps. Target preflights, positive batches, and
+negative singletons share a quota-aware worker pool capped at 32 by default;
+`--jobs` is the explicit override for constrained or dedicated hosts. An
+expected deterministic crash or timeout is not rerun serially; changed,
+unclassified, and otherwise transient-looking results are.
+
 The migration preserves the former pytest item's result mapping:
 
 | Condition | CTest result |