fix(model): isolate observed-overlap priors

CMakeLists.txt

@@ -125,6 +125,7 @@ if(BUILD_TESTING)
   # 跨进程学习与故障隔离。 / Cross-process learning and failure isolation.
   add_executable(learning_integration_tests tests/learning_integration_tests.cpp)
   target_link_libraries(learning_integration_tests PRIVATE same_core)
+  target_compile_definitions(learning_integration_tests PRIVATE SAME_TEST_ROOT="${CMAKE_CURRENT_SOURCE_DIR}")
   if(WIN32)
     target_link_libraries(learning_integration_tests PRIVATE shell32)
   endif()

docs/contextual-learning.md

@@ -18,6 +18,8 @@ predicted_ms = xᵀ β_backend
 四项分别表示固定成本、字节成本、批量提交成本和与 payload 交互的竞争成本。CPU/iGPU 的 C 包括共享主机资源的在途任务；CUDA 使用同后端任务数。C 是轻量代理而非带宽测量。误差诊断始终比较**选择时冻结的预测**与结果，拟合则使用实际观测到的重叠上下文，避免把事后信息误称为决策时已知。设备的名称、厂商、驱动、实现版本、架构、计算单元、硬件线程、内存约束、统一内存属性和实际批量参与设备身份键；它们**不是全部作为回归变量输入**。因此这是本机按设备拟合，不是跨任意 GPU 的零样本性能预测。
 The four terms represent fixed, byte, batch and payload–contention costs. Decision prediction uses the selection-time context; fitting uses observed overlap, while the error diagnostic stays tied to the actual decision-time prediction. Contention is still a proxy, not a bandwidth meter or a causal estimate. Device characteristics namespace learned state; this does not promise zero-shot transfer.
 
+旧版本把选择时的在途数用于拟合，新版本则使用任务期间观测到的重叠；虽然四维公式未变，样本特征的**语义**已经变化。新模型身份为 `same-learning-key-v2`，不会把 v0.6.0 的 `v1` 统计与新标签混用。旧数据库与来源账本保留，不清空用户状态；第一次新版本运行需要重新积累或执行 `train`，后续相同工作区、设备与配置仍会正常加载 v2 先验。summary 的 `Model identity` 行可用于识别这一有意的冷启动。/ The feature formula is unchanged but its contention-label semantics changed. `same-learning-key-v2` isolates new observations from v0.6.0 v1 state without deleting old rows. The first new run must relearn or retrain; subsequent matching runs load v2 priors.
+
 ## 数学模型 / Mathematical model
 
 每后端维护可加充分统计量（sufficient statistics）：S=Σxxᵀ、t=Σxy、q=Σy² 和权重 W，以及特征范围。固定 4×4 数组，无逐样本历史分配。前四个本轮样本每次拟合，之后每 16 个样本重拟合，预测使用缓存系数。求解采用对角尺度归一化的岭回归（ridge regression）：

docs/igpu-worker-concurrency.md

@@ -75,6 +75,12 @@ uses the decision-time prediction, so it does not benefit from information that
 became available only after dispatch. This remains a coarse contention proxy;
 overlap duration and DRAM traffic are not observed.
 
+The fitting-context change is not compatible with old v0.6.0 learned
+statistics even though the four algebraic features have the same shape. The
+application uses `same-learning-key-v2` for all three backends, preserving old
+database rows while preventing mixed-semantic priors. This intentionally
+introduces one cold-learning run after upgrade; later runs reuse v2 evidence.
+
 ## Local validation evidence
 
 Windows x64 MSVC 19.51 CPU-only Debug CTest passed 27/27 after the

src/application.cpp

@@ -191,6 +191,9 @@ void prepare_state(const fs::path& root) {
 }
 /// 运行边界遗忘，不随工作线程数变化。 / Run-boundary forgetting, independent of worker count.
 constexpr double learning_decay = 0.9;
+/// 新标签使用实际重叠上下文，必须与旧的选择时标签分开存放。
+/// Observed-overlap labels must not mix with prior selection-time labels.
+constexpr std::string_view learning_identity = "same-learning-key-v2";
 /// 长度前缀消除字段拼接歧义，超长身份拒绝学习而不截断碰撞。
 /// Length prefixes avoid ambiguous identities; reject oversize keys rather than truncate.
 std::string learning_key(const Config& config, const DeviceProfile& cpu, BackendKind backend,
@@ -216,7 +219,7 @@ std::string learning_key(const Config& config, const DeviceProfile& cpu, Backend
         number(value.effective_batch_bytes);
         number(value.unified_memory);
     };
-    add("same-learning-key-v1");
+    add(learning_identity);
     number(detail::OnlineModel::feature_version);
     number(detail::OnlineModel::feature_count);
     add(backend_name(backend));
@@ -413,6 +416,8 @@ void render_learning(const Learning& learning, bool enabled, std::ostream& out,
         report.row("Persistence", learning.store ? "enabled" : "disabled",
                    learning.store ? ReportTone::good : ReportTone::warning);
         report.row("Model database", ".same/model.db", ReportTone::info);
+        report.row("Model identity", std::string(learning_identity) + " | observed overlap",
+                   ReportTone::info);
         report.row("Disabled reason",
                    learning.disabled_reason.empty() ? "none" : learning.disabled_reason,
                    learning.disabled_reason.empty() ? ReportTone::muted : ReportTone::warning);
@@ -442,6 +447,7 @@ void render_learning(const Learning& learning, bool enabled, std::ostream& out,
         << " model_persistence_disabled_reason="
         << (learning.disabled_reason.empty() ? "none" : learning.disabled_reason)
         << " model_db=.same/model.db"
+        << " model_identity=" << learning_identity
         << " model_load_attempts=" << learning.loads.load()
         << " model_prior_hits=" << learning.hits.load()
         << " model_invalid=" << learning.invalid.load() << " model_saved_keys=" << learning.saved

tests/learning_integration_tests.cpp

@@ -184,6 +184,24 @@ std::string stored_key(const fs::path& path) {
     sqlite3_close(db);
     return key;
 }
+/// 测试升级场景时仅移除当前身份，保留旧版来源贡献。
+/// Remove only the current identity while retaining the legacy contribution for upgrade tests.
+void remove_key(const fs::path& path, std::string_view key) {
+    sqlite3* db{};
+    check(sqlite3_open(utf8(path).c_str(), &db) == SQLITE_OK, "open model for identity test");
+    for (const auto* table : {"models", "model_parts"}) {
+        const auto sql = std::string("DELETE FROM ") + table + " WHERE key=?1";
+        sqlite3_stmt* statement{};
+        check(sqlite3_prepare_v2(db, sql.c_str(), -1, &statement, nullptr) == SQLITE_OK,
+              "prepare identity removal");
+        check(sqlite3_bind_blob(statement, 1, key.data(), static_cast<int>(key.size()),
+                                SQLITE_TRANSIENT) == SQLITE_OK &&
+                  sqlite3_step(statement) == SQLITE_DONE,
+              "remove current identity");
+        sqlite3_finalize(statement);
+    }
+    sqlite3_close(db);
+}
 /// 汇总中的明确计数，不解析展示顺序。 / Extract an explicit summary count independent of ordering.
 std::uint64_t metric(const std::string& text, const std::string& name) {
     std::smatch match;
@@ -197,7 +215,7 @@ int main(int argc, char** argv) {
     try {
         check(argc == 2, "expected CLI executable");
         const auto exe = executable_path(argv[1]);
-        base = fs::canonical(fs::temp_directory_path()) /
+        base = fs::canonical(fs::path(SAME_TEST_ROOT)) / ".temp" /
                ("same-learning-" +
                 std::to_string(std::chrono::steady_clock::now().time_since_epoch().count()));
         const auto root = base / "root";
@@ -263,6 +281,34 @@ int main(int argc, char** argv) {
             check(std::abs(second[0].weight - (0.9 * first[0].weight + 16)) < 1e-8,
                   "decayed prior not merged exactly once");
         }
+        // v0.6.0 的旧特征标签不能被新的实际重叠语义复用，但数据库须保持可写。
+        // Old selection-time labels must not load as observed-overlap priors; DB remains writable.
+        const auto current_prefix = std::string("same-learning-key-v2");
+        const auto prefix_at = key.find(current_prefix);
+        check(prefix_at != std::string::npos, "new model identity namespace missing");
+        auto legacy_key = key;
+        legacy_key.replace(prefix_at, current_prefix.size(), "same-learning-key-v1");
+        same::detail::OnlineModel::State legacy{};
+        {
+            same::ModelStore store(database);
+            legacy = *store.load(key);
+            check(store.save(legacy_key, legacy), "cannot seed legacy model contribution");
+        }
+        remove_key(database, key);
+        run();
+        check(metric(read(err), "model_prior_hits") == 0 &&
+                  metric(read(err), "model_persistence_enabled") == 1 && read(out) == expected,
+              "legacy model identity leaked into new routing or disabled persistence");
+        {
+            same::ModelStore store(database);
+            check(store.load(legacy_key) &&
+                      store.load(legacy_key)->at(0).samples == legacy[0].samples &&
+                      store.load(key) && store.load(key)->at(0).samples == 16,
+                  "new training erased legacy state or failed to create isolated state");
+        }
+        run();
+        check(metric(read(err), "model_prior_hits") == 2,
+              "new identity did not resume normal cross-run reuse");
         const auto before = read(database);
         run(false);
         check(read(database) == before, "no-pgo changed model database");