Implement manifest-driven remote cloud backup transport

.github/WORKFLOW_GUIDELINES.md

@@ -4,7 +4,7 @@
 Diese Richtlinie gilt fuer den schlanken, release-zentrierten Workflow-Kern.
 Die kanonische Liste aktiver Workflows steht in `.github/WORKFLOW_REGISTRY.md`.
 
-## Aktive Workflows (43)
+## Aktive Workflows (44)
 Die aktuelle kanonische Liste steht in `.github/WORKFLOW_REGISTRY.md`; der alte 21er-Stand war veraltet und wird hier durch den aktuellen, im Repository geltenden Zustand ersetzt.
 
 Kernliste der aktiven Workflows:
@@ -28,6 +28,7 @@ Kernliste der aktiven Workflows:
 - `.github/workflows/compliance-supply-chain.yml`
 - `.github/workflows/build-ollama-router.yml`
 - `.github/workflows/gate-copilot-regression.yml`
+- `.github/workflows/copilot-code-review.yml`
 - `.github/workflows/publish-wiki.yml`
 - `.github/workflows/release-docker-image.yml`
 - `.github/workflows/edition-hyperscaler-ci.yml`
@@ -90,6 +91,9 @@ Kernliste der aktiven Workflows:
 - `concurrency` mit `cancel-in-progress` auf Push/PR-Workflows setzen.
 - Berechtigungen minimal halten (`permissions` least privilege).
 - Schwere Benchmark-, GPU- und Sweep-Jobs standardmaessig ueber `schedule` oder `workflow_dispatch` isolieren.
+- Copilot-Review-Runner-Konfigurationen muessen als selbstbegrenzte Workflows
+  mit dem Jobnamen `copilot-setup-steps` und einem expliziten Ubuntu
+  `runs-on` deklariert werden.
 
 ## Security Guidelines
 - Keine Secrets im YAML oder in Shell-Skripten hardcoden.

.github/WORKFLOW_REGISTRY.md

@@ -64,6 +64,8 @@ Signalqualität und Release-Stabilitaet zu verbessern.
   — Scoped CI fuer `tools/copilot-ollama-router/**`
 - `.github/workflows/gate-copilot-regression.yml`
   — Copilot/CMake-Regression Guard
+- `.github/workflows/copilot-code-review.yml`
+  — Self-scoped Copilot review runner declaration (`copilot-setup-steps`) so agentic reviews have an assigned Ubuntu runner
 - `.github/workflows/publish-wiki.yml`
   — Publishes docs/architecture, docs/governance, src/*/ROADMAP.md and developer wiki to GitHub Wiki on push to develop or manual dispatch; community guardrail blocks private plugin paths
 
@@ -117,7 +119,7 @@ Geplante Dateinamen-Harmonisierung (Soll-Format aus Workflow-Design):
 - `.github/docs/WORKFLOW_FILENAME_RENAME_MATRIX.md`
 
 ## Stand
-- Aktive Workflows im Verzeichnis `.github/workflows/`: 43
+- Aktive Workflows im Verzeichnis `.github/workflows/`: 44
 - Deaktivierte Workflows in `.github/no_workflows/`: 30
 - Strategie: Lean + harte Triggergrenzen + Quarantaene fuer uebertriggernde CI
 - Der 21er-Zähler war im vorherigen Dokumentationsstand veraltet; der aktuelle Stand wird durch die kanonische Liste in diesem Registry-Dokument und die zugehörigen Workflow-Dateien definiert.

.github/workflows/build-clang-fast.yml

@@ -42,7 +42,7 @@ jobs:
       submodules: false
       working_directory: .
       setup_command: >-
-        set -euo pipefail; sudo apt-get update -qq; sudo apt-get install -y --no-install-recommends clang lld cmake ninja-build pkg-config librocksdb-dev libgtest-dev libssl-dev zlib1g-dev libzstd-dev libfmt-dev libspdlog-dev nlohmann-json3-dev libtbb-dev libyaml-cpp-dev libmimalloc-dev libcurl4-openssl-dev libvulkan-dev glslc libcpp-httplib-dev libboost-system-dev libboost-filesystem-dev libgrpc++-dev libprotobuf-dev protobuf-compiler-grpc libpugixml-dev
+        set -euo pipefail; sudo apt-get update -qq; sudo apt-get install -y --no-install-recommends clang lld cmake ninja-build pkg-config sccache librocksdb-dev libgtest-dev libssl-dev zlib1g-dev libzstd-dev libfmt-dev libspdlog-dev nlohmann-json3-dev libtbb-dev libyaml-cpp-dev libmimalloc-dev libcurl4-openssl-dev libvulkan-dev glslc libcpp-httplib-dev libboost-system-dev libboost-filesystem-dev libgrpc++-dev libprotobuf-dev protobuf-compiler-grpc libpugixml-dev
       configure_command: >-
         cmake -S . -B build-clang-fast -G Ninja -DCMAKE_BUILD_TYPE=RelWithDebInfo -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_CXX_FLAGS="-Wall -Wextra -Wpedantic" -DTHEMIS_BUILD_BENCHMARKS=OFF -DTHEMIS_BUILD_TESTS=ON
       build_command: >-

.github/workflows/copilot-code-review.yml

@@ -0,0 +1,32 @@
+name: Copilot code review setup
+
+on:
+  workflow_dispatch:
+  push:
+    branches: [develop, community]
+    paths:
+      - '.github/workflows/copilot-code-review.yml'
+  pull_request:
+    branches: [develop, community]
+    paths:
+      - '.github/workflows/copilot-code-review.yml'
+
+permissions:
+  contents: read
+
+concurrency:
+  group: copilot-code-review-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  # The job name must remain `copilot-setup-steps` so Copilot can detect it.
+  copilot-setup-steps:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1
+        with:
+          persist-credentials: false
+
+      - name: Confirm runner availability
+        run: echo "Copilot code review runner is configured."

.gitmodules

@@ -48,3 +48,9 @@
 [submodule "projects/Themis.AdminTools.Shared"]
 	path = projects/Themis.AdminTools.Shared
 	url = https://github.com/makr-code/themisdb_admin_tools.git
+
+[submodule "plugins/themisdb_ethic_ai"]
+	path = plugins/themisdb_ethic_ai
+	url = https://github.com/makr-code/themisdb_ethic_ai.git
+	branch = develop
+	commit = ce401ad9d604012a2c02655e79f5c17f57a9f82d

ROADMAP.md

@@ -36,14 +36,15 @@ ThemisDB is a high-performance multi-model database with native AI/LLM integrati
 |---|---|---|---|
 | core | 9 listed | Mostly DOC / evidence gaps | Runtime adapter registry and plugin loading are delivered; remaining items are Wave D operability and refreshed evidence |
 | base | 8 listed | Mostly historical scanner noise | `src/base/MODULE_GAPS.md` re-scan shows 0 actionable current gaps; remaining items are documented false positives or follow-up docs |
-| server | 4 residual source gaps | REAL IMPL gaps | gRPC-Web proxy is still UNIMPLEMENTED-only in fallback builds; time-series aggregate/retention providers and RoPE metrics still rely on degraded fallback paths; non-Linux MCP stdio transport remains unsupported |
+| server | 1 residual source gap | MOSTLY REMEDIATED | gRPC-Web fallback builds now advertise an explicit fail-closed capability contract, RoPE DELETE now disables runtime config, and unsupported MCP stdio self-disables; remaining work is first-class time-series provider DI beyond degraded metadata signaling |
 | query | residual perf / validation follow-up | Mostly verification / perf gaps | Process-mining trace/pattern/ideal query paths and the three `ETHICS_*` runtime gaps were closed on 2026-08-31; remaining work is optimizer/federation hardening and benchmark evidence |
 | transaction | 19 listed | Mostly verification / benchmark evidence | Wave 4C code gaps are closed; build/run, chaos, and representative-hardware evidence remain open |
 | auth | 8+ listed | Mostly verification / perf follow-up | Wave 4B source gaps are closed; remaining work is Wave 8 tests, representative-hardware baselines, and protocol-matrix regressions |
 | LLM | 13 listed | MIXED | Major Wave 5 closures landed; remaining real gaps center on distributed collectives, multi-tenant isolation, and final cross-module speculative/TARG wiring |
 | RAG / LLM Wiki | 57 listed | Mostly perf / integration follow-up | BM25+, RRF, persistent cache, and real `LLMJudgeIntegration` path are implemented; remaining work is performance gates, Recall@k sign-off, Wikipedia ABI wiring, and entropy-bridge integration |
-| GPU/CUDA | 21+53 listed | REAL IMPL gaps | CUDA/HIP kernel parity and representative-hardware validation remain open release blockers for acceleration-heavy paths |
-| access_model | 21 listed | REAL IMPL gaps | Benchmarks and GATE-ACM-01..06 not yet implemented |
+| GPU/CUDA | 21+53 listed | REAL IMPL gaps | Break-even routing now uses production build wiring plus explicit CPU/GPU profiling contracts, but CUDA/HIP kernel parity, unchecked-kernel-call closure, and representative-hardware validation remain open release blockers |
+| storage | 11 real gaps after 2026-08-31 revalidation | MIXED | Backup restore fail-closed hardening, ggml bridge runtime wiring, `SecuritySignatureManager` null-backend fail-closed behavior, and remote S3/GCS/Azure manifest transport are in place; biggest remaining gaps are long-run validation evidence and cloud-backend hardening follow-up |
+| access_model | roadmap contradiction | Mostly DOC drift | Source and module evidence show Phase 5-6 observability, e2e/concurrency tests, and GATE-ACM-01..06 are complete; stale checklist/known-issues text must stay synchronized |
 
 ## Release Hardening Program (current canonical version: v2.4.0-alpha)
 
@@ -71,7 +72,7 @@ ThemisDB is a high-performance multi-model database with native AI/LLM integrati
 - [x] Wave 8, chaos/fault-injection, sanitizer/recovery, penetration-test, and 99.99% SLA sign-off artefacts are closed: sanitizer evidence bundle at `docs/security/GA_SANITIZER_EVIDENCE_BUNDLE.md`; pentest evidence bundle at `security/pentest/GA_PENTEST_EVIDENCE_BUNDLE.md`; Wave 9 SLA/chaos gates PASS; final governance sign-off pending human approval at `docs/governance/GA_PROMOTION_SIGN_OFF.md`.
 - [x] Phase 1-6 execution contract complete: all technical gates PASS; human sign-off (Section 9 of `docs/governance/GA_PROMOTION_SIGN_OFF.md`) is the only remaining GA blocker.
 - [x] Tools build-option transition complete: canonical flag for desktop tools is `THEMIS_BUILD_TOOLS` (default `ON`); legacy alias removed.
-- [~] Core-first residual source-gap queue revalidated: finish server runtime fallback gaps and query feature gaps first, then close LLM/RAG integration and GPU representative-hardware gates (Target: Q4 2026).
+- [~] Core-first residual source-gap queue revalidated: finish the remaining server time-series provider DI gap and query feature gaps first, then close LLM/RAG integration and the remaining GPU parity / representative-hardware gates after the 2026-08-31 acceleration break-even hardening batch. (Target: Q4 2026).
 
 ## Program Execution Model (Wave A → B → C → D)
 

cmake/CMakeLists.txt

@@ -2909,6 +2909,7 @@ set(THEMIS_CORE_SOURCES
     ../src/acceleration/geo_acceleration_bridge.cpp
     ../src/acceleration/device_manager.cpp
     ../src/acceleration/vllm_resource_manager.cpp
+    ../src/acceleration/break_even_validator.cc
     # PERF-D3: Parallel batch insertion + SIMD distance pipeline
     ../src/acceleration/vec_knn.cpp
     # shader_integrity.cpp uses only OpenSSL (SHA-256) — no GPU/graphics dependency.

cmake/ModularBuild.cmake

@@ -309,6 +309,7 @@ set(THEMIS_BASE_SOURCES
     ../src/acceleration/plugin_security.cpp
     ../src/acceleration/device_manager.cpp
     ../src/acceleration/vllm_resource_manager.cpp
+    ../src/acceleration/break_even_validator.cc
     ../src/acceleration/shader_integrity.cpp
     # PERF-D3: Parallel batch insertion + SIMD distance pipeline
     ../src/acceleration/vec_knn.cpp

include/acceleration/break_even_validator.h

@@ -15,6 +15,7 @@
 
 #include <string>
 #include <chrono>
+#include <functional>
 #include <optional>
 #include <unordered_map>
 #include <mutex>
@@ -124,10 +125,10 @@ struct BreakEvenDecision {
     float speedup_ratio = 0.0f;
 
     /// CPU path execution time (milliseconds)
-    std::chrono::milliseconds cpu_time_ms;
+    std::chrono::milliseconds cpu_time_ms{0};
 
     /// GPU path execution time including transfer (milliseconds)
-    std::chrono::milliseconds gpu_time_ms;
+    std::chrono::milliseconds gpu_time_ms{0};
 
     /// Human-readable reason: "break_even_met", "gpu_unavailable", etc.
     std::string reason;
@@ -170,6 +171,11 @@ struct BreakEvenDecision {
  */
 class BreakEvenValidator {
 public:
+    using ProfileFn = std::function<std::optional<std::chrono::milliseconds>(
+        const WorkloadProfile&)>;
+    using MetricsSinkFn = std::function<void(
+        const WorkloadProfile&, const BreakEvenDecision&)>;
+
     /**
      * @brief Construct a new BreakEvenValidator instance.
      *
@@ -197,7 +203,7 @@ class BreakEvenValidator {
      * @param profile Workload profile defining input characteristics
      * @return BreakEvenDecision with recommendation and metrics
      *
-     * @thread Fully thread-safe; protected by internal mutex
+     * @note Thread safety: Fully thread-safe; protected by internal mutex
      */
     BreakEvenDecision ShouldUseGPU(const WorkloadProfile& profile);
 
@@ -213,7 +219,7 @@ class BreakEvenValidator {
      * @param profile Workload profile to profile
      * @return BreakEvenDecision with fresh profiling results
      *
-     * @thread Fully thread-safe
+     * @note Thread safety: Fully thread-safe
      */
     BreakEvenDecision Profile(const WorkloadProfile& profile);
 
@@ -226,7 +232,7 @@ class BreakEvenValidator {
      * @param kernel Kernel type to set threshold for
      * @param threshold Minimum speedup ratio (must be >= 1.0)
      *
-     * @thread Fully thread-safe
+     * @note Thread safety: Fully thread-safe
      */
     void SetSpeedupThreshold(KernelType kernel, float threshold);
 
@@ -236,7 +242,7 @@ class BreakEvenValidator {
      * @param kernel Kernel type
      * @return Current threshold for this kernel (or default 1.5 if not set)
      *
-     * @thread Fully thread-safe
+     * @note Thread safety: Fully thread-safe
      */
     float GetSpeedupThreshold(KernelType kernel) const;
 
@@ -245,7 +251,7 @@ class BreakEvenValidator {
      *
      * Next call to ShouldUseGPU() will trigger profiling (cache miss).
      *
-     * @thread Fully thread-safe
+     * @note Thread safety: Fully thread-safe
      */
     void ClearCache();
 
@@ -257,10 +263,51 @@ class BreakEvenValidator {
      *
      * @param duration Cache validity duration
      *
-     * @thread Fully thread-safe
+     * @note Thread safety: Fully thread-safe
      */
     void SetCacheValidityDuration(std::chrono::hours duration);
 
+    /**
+     * @brief Override CPU workload profiling with a caller-provided implementation.
+     *
+     * When set, ProfileCPU() delegates to @p fn instead of the built-in
+     * deterministic cost model. Passing an empty function restores the default
+     * CPU estimator. Any cached decisions are cleared so subsequent calls are
+     * re-profiled with the new behavior.
+     *
+     * @param fn CPU profiling callback, or empty to restore defaults
+     *
+     * @note Thread safety: Fully thread-safe
+     */
+    void SetCPUProfileFn(ProfileFn fn);
+
+    /**
+     * @brief Override GPU workload profiling with a caller-provided implementation.
+     *
+     * When set, ProfileGPU() delegates to @p fn instead of the built-in
+     * deterministic GPU estimate. Passing an empty function restores the default
+     * GPU estimator. Any cached decisions are cleared so subsequent calls are
+     * re-profiled with the new behavior.
+     *
+     * @param fn GPU profiling callback, or empty to restore defaults
+     *
+     * @note Thread safety: Fully thread-safe
+     */
+    void SetGPUProfileFn(ProfileFn fn);
+
+    /**
+     * @brief Register a metrics sink that receives every fresh profile decision.
+     *
+     * The sink is called after a non-cached profile completes successfully or
+     * degrades to CPU because GPU profiling is unavailable. Exceptions thrown by
+     * the sink are swallowed to preserve fail-closed decision behavior.
+     *
+     * @param fn Metrics callback, or empty to disable metrics emission
+     *
+     * @note Thread safety: Fully thread-safe
+     */
+    void SetMetricsSink(MetricsSinkFn fn);
+
     /**
      * @brief Get the latest break-even speedup ratio for a kernel type.
      *
@@ -270,7 +317,7 @@ class BreakEvenValidator {
      * @param kernel Kernel type to query
      * @return Latest observed speedup ratio, or 0.0 if no data
      *
-     * @thread Fully thread-safe
+     * @note Thread safety: Fully thread-safe
      */
     float GetLatestBreakEvenRatio(KernelType kernel) const;
 
@@ -279,7 +326,7 @@ class BreakEvenValidator {
      *
      * @return Number of ShouldUseGPU() calls that hit the cache
      *
-     * @thread Fully thread-safe
+     * @note Thread safety: Fully thread-safe
      */
     size_t GetCacheHitCount() const;
 
@@ -288,7 +335,7 @@ class BreakEvenValidator {
      *
      * @return Number of ShouldUseGPU() calls that missed the cache
      *
-     * @thread Fully thread-safe
+     * @note Thread safety: Fully thread-safe
      */
     size_t GetCacheMissCount() const;
 
@@ -297,7 +344,7 @@ class BreakEvenValidator {
      *
      * @return Number of cached decision entries
      *
-     * @thread Fully thread-safe
+     * @note Thread safety: Fully thread-safe
      */
     size_t GetCacheSize() const;
 
@@ -321,8 +368,8 @@ class BreakEvenValidator {
     /**
      * @brief Profile CPU execution time for the given workload.
      *
-     * Delegates to CPU reference kernel implementations to measure
-     * end-to-end execution time (no GPU transfer overhead).
+     * Uses a caller-provided profiling hook when configured; otherwise falls
+     * back to the built-in deterministic CPU cost model.
      *
      * @param profile Workload to profile on CPU
      * @return CPU execution time, or nullopt if profiling failed
@@ -333,16 +380,22 @@ class BreakEvenValidator {
     /**
      * @brief Profile GPU execution time for the given workload.
      *
-     * Includes GPU allocation, data transfer, kernel execution, and sync time.
-     * Returns nullopt if GPU is unavailable or profiling fails.
+     * Uses a caller-provided profiling hook when configured; otherwise falls
+     * back to the built-in deterministic GPU estimate including transfer and
+     * launch overhead. Returns nullopt if the selected device cannot run GPU
+     * work or profiling fails.
      *
      * @param profile Workload to profile on GPU
      * @return GPU execution time (transfer + compute), or nullopt if GPU unavailable
      */
     std::optional<std::chrono::milliseconds> ProfileGPU(
         const WorkloadProfile& profile);
 
-
+    static bool RequiresVectorDimension(KernelType kernel);
+    static bool IsGpuCapableDevice(DeviceType device);
+    static double EstimateWorkUnits(const WorkloadProfile& profile);
+    static std::optional<std::chrono::milliseconds> MillisecondsFromEstimate(
+        double estimated_ms);
 
     /**
      * @brief Parse KernelType from string.
@@ -394,6 +447,11 @@ class BreakEvenValidator {
 
     // Cache configuration
     std::chrono::hours cache_validity_duration_;
+
+    // Optional profiling/metrics hooks
+    ProfileFn cpu_profile_fn_;
+    ProfileFn gpu_profile_fn_;
+    MetricsSinkFn metrics_sink_;
 };
 
 }  // namespace acceleration