gfx12: STRATA_HEAD_MIX_MULTI=1 by default (exact; own function, droppable alone)

The final hyper-connection mixer of the verify window and of the draft layer as one fused_gr_read_multi instead of one gr_read per
token (PR 3264, author bsorensen110).  Kept apart from the main gfx12 table in gfx12_extra.hpp so it merges with the table on
another branch without touching it; STRATA_GFX12_DEFAULTS=0, STRATA_DEFAULTS=026 and a variable the user sets decide first.
R9700 (p620, IQ3_S, serve, static residency, 8 interleaved pairs with an A/A arm): json_extract +0.82% (6/8), agent +0.63% (6/8),
code_py -0.96% (3/8, that prompt's A/A moves 3.5% a pair); greedy ids identical in 16/16 and 15/16 request sets.

CMakeLists.txt

@@ -1826,6 +1826,9 @@ if(STRATA_BUILD_TESTS)
   target_include_directories(amd_community_test_dot_off PRIVATE include)
   target_compile_definitions(amd_community_test_dot_off PRIVATE EXPECT_DOT_BUILT=0 STRATA_AMD_COMMUNITY_DOT_OFF=1)
   add_test(NAME amd_community_test_dot_off COMMAND amd_community_test_dot_off WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR})
+  add_executable(gfx12_extra_test src/core/gfx12_extra_test.cpp src/core/arch_defaults.cpp)   # gfx12 extra defaults and their off switches
+  target_include_directories(gfx12_extra_test PRIVATE include)
+  add_test(NAME gfx12_extra_test COMMAND gfx12_extra_test)
   add_executable(nv_defaults_test src/core/nv_defaults_test.cpp src/core/nv_defaults.cpp)   # the sm_121 (GB10) table and its two off switches
   target_include_directories(nv_defaults_test PRIVATE include)
   add_test(NAME nv_defaults_test COMMAND nv_defaults_test)

include/strata/core/gfx12_extra.hpp

@@ -0,0 +1,36 @@
+// include/strata/core/gfx12_extra.hpp - defaults for gfx12 (RDNA4: R9700, RX 9070, RX 9060) that came out of the PR-AMD pass of 2026-11-01 and are
+// kept apart from the main table (arch_defaults.cpp) so each can be dropped alone.  Same off switches as that table: STRATA_GFX12_DEFAULTS=0
+// and STRATA_DEFAULTS=026 give the old behaviour, and a variable the user sets is never overridden.
+//
+// Exact (bit for bit the same answers):
+//   STRATA_HEAD_MIX_MULTI=1  the final hyper-connection mixer of the verify window and of the draft layer as one fused_gr_read_multi
+//                            instead of one gr_read per token (#3264, bsorensen110).  R9700 (p620, IQ3_S, serve, static residency,
+//                            8 interleaved pairs, A/A beside it): json_extract +0.82% (6/8 pairs), agent +0.63% (6/8), code_py -0.96%
+//                            (3/8; that prompt's A/A moves +-3.5% a pair), greedy ids identical in 16/16 and 15/16 request sets.
+#pragma once
+
+#include <cstdlib>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "strata/core/defaults_026.hpp"
+#include "strata/kernels/gfx_arch.hpp"
+
+namespace strata::core {
+
+/// `defaults026`: STRATA_DEFAULTS=026 (defaults_026.hpp reads it once, so a test passes it).
+inline std::vector<std::pair<std::string, std::string>> gfx12_exact_extra_env(const char* gcn_arch, bool defaults026) {
+    std::vector<std::pair<std::string, std::string>> t;
+    if (!strata::kernels::gfx_arch_is_gfx12(gcn_arch)) return t;
+    const char* off = std::getenv("STRATA_GFX12_DEFAULTS");
+    if (off != nullptr && off[0] == '0') return t;
+    if (defaults026) return t;
+    t.push_back({"STRATA_HEAD_MIX_MULTI", "1"});
+    return t;
+}
+inline std::vector<std::pair<std::string, std::string>> gfx12_exact_extra_env(const char* gcn_arch) {
+    return gfx12_exact_extra_env(gcn_arch, strata::qwen_defaults_026());
+}
+
+}  // namespace strata::core

src/core/arch_defaults.cpp

@@ -2,6 +2,7 @@
 #include "strata/core/arch_defaults.hpp"
 #include "strata/core/defaults_026.hpp"
 #include "strata/core/amd_community.hpp"
+#include "strata/core/gfx12_extra.hpp"
 
 #include "strata/kernels/gfx_arch.hpp"
 
@@ -208,6 +209,19 @@ std::vector<std::string> apply_arch_defaults(const char* gcn_arch, bool integrat
         set_default(kv);
         set.push_back(kv.first);
     }
+    {   // gfx12 extras kept apart from the table (gfx12_extra.hpp); not counted in the returned list
+        std::vector<std::string> extra;
+        for (const auto& kv : gfx12_exact_extra_env(gcn_arch)) {
+            if (std::getenv(kv.first.c_str()) != nullptr) continue;
+            set_default(kv);
+            extra.push_back(kv.first);
+        }
+        if (!extra.empty()) {
+            std::fprintf(stderr, "strata: gfx12 (RDNA4): exact extra switches on by default (STRATA_GFX12_DEFAULTS=0 turns them off; a switch you set is kept): ");
+            for (size_t i = 0; i < extra.size(); ++i) std::fprintf(stderr, "%s%s", i ? " " : "", extra[i].c_str() + 7);
+            std::fprintf(stderr, "\n");
+        }
+    }
     if (!set.empty()) {
         const bool g12 = strata::kernels::gfx_arch_is_gfx12(gcn_arch);
         std::fprintf(stderr, "strata: %s: %zu exact speed switches on by default (%s=0 turns "

src/core/gfx12_extra_test.cpp

@@ -0,0 +1,61 @@
+// src/core/gfx12_extra_test.cpp - the gfx12 extra defaults (gfx12_extra.hpp) and their off switches (no GPU needed).
+#include "strata/core/arch_defaults.hpp"
+#include "strata/core/gfx12_extra.hpp"
+
+#include <cstdio>
+#include <cstdlib>
+#include <string>
+
+static int fails = 0;
+#define CHECK(c) do { if (!(c)) { std::printf("FAIL line %d: %s\n", __LINE__, #c); ++fails; } } while (0)
+static void unset(const char* k) {
+#if defined(_WIN32)
+    _putenv_s(k, "");
+#else
+    unsetenv(k);
+#endif
+}
+static void put(const char* k, const char* v) {
+#if defined(_WIN32)
+    _putenv_s(k, v);
+#else
+    setenv(k, v, 1);
+#endif
+}
+static bool has(const std::vector<std::pair<std::string, std::string>>& t, const char* k, const char* v) {
+    for (const auto& kv : t) if (kv.first == k && kv.second == v) return true;
+    return false;
+}
+
+int main() {
+    using namespace strata::core;
+    unset("STRATA_GFX12_DEFAULTS");
+    unset("STRATA_DEFAULTS");
+    unset("STRATA_HEAD_MIX_MULTI");
+    // the exact extra: gfx12 only
+    CHECK(has(gfx12_exact_extra_env("gfx1201"), "STRATA_HEAD_MIX_MULTI", "1"));
+    CHECK(has(gfx12_exact_extra_env("gfx1200:xnack-"), "STRATA_HEAD_MIX_MULTI", "1"));
+    for (const char* other : {"gfx1100", "gfx1151", "gfx1030", "gfx906", "gfx1010", "", "gfx12010"}) CHECK(gfx12_exact_extra_env(other).empty());
+    CHECK(gfx12_exact_extra_env(nullptr).empty());
+    // the off switches
+    put("STRATA_GFX12_DEFAULTS", "0");
+    CHECK(gfx12_exact_extra_env("gfx1201").empty());
+    unset("STRATA_GFX12_DEFAULTS");
+    CHECK(gfx12_exact_extra_env("gfx1201", true).empty());     // STRATA_DEFAULTS=026
+    CHECK(!gfx12_exact_extra_env("gfx1201", false).empty());
+    // apply: sets it, keeps a value the user set, does nothing off gfx12 or with the switch off
+    apply_arch_defaults("gfx1201", false);
+    CHECK(std::getenv("STRATA_HEAD_MIX_MULTI") != nullptr && std::string(std::getenv("STRATA_HEAD_MIX_MULTI")) == "1");
+    unset("STRATA_HEAD_MIX_MULTI");
+    put("STRATA_HEAD_MIX_MULTI", "0");
+    apply_arch_defaults("gfx1201", false);
+    CHECK(std::string(std::getenv("STRATA_HEAD_MIX_MULTI")) == "0");
+    unset("STRATA_HEAD_MIX_MULTI");
+    apply_arch_defaults("gfx1100", false);
+    CHECK(std::getenv("STRATA_HEAD_MIX_MULTI") == nullptr);
+    put("STRATA_GFX12_DEFAULTS", "0");
+    apply_arch_defaults("gfx1201", false);
+    CHECK(std::getenv("STRATA_HEAD_MIX_MULTI") == nullptr);
+    if (fails == 0) std::printf("gfx12_extra_test: OK\n");
+    return fails == 0 ? 0 : 1;
+}