Edgeguard implementation

AGENTS.md

@@ -695,3 +695,66 @@ Entry format:
 - Details: `ThHfModelBase` keeps Transformers/PT as the default GPU and fallback path, but CPU-only `HF_RUNTIME=auto` now loads `artifact_manifest.json`, selects a declared ONNX Runtime artifact, downloads only safe allow-patterns, loads schema and contract decoder from HF artifacts, and exposes the decoded artifact contract through the existing text-classifier flow. Business API response shaping now passes through generic model/runtime metadata emitted by serving.
 - Verification: `python3 -m unittest extensions.serving.test_th_hf_model_base extensions.serving.test_th_text_classifier extensions.serving.test_th_privacy_filter extensions.business.edge_inference_api.test_text_classifier_inference_api extensions.business.edge_inference_api.test_privacy_filter_inference_api`; `python3 -m py_compile extensions/serving/default_inference/nlp/th_hf_model_base.py extensions/business/edge_inference_api/text_classifier_inference_api.py`; required serving gate `python3 -m unittest extensions.serving.model_testing.test_llm_servings` currently fails at import with `ImportError: cannot import name 'Logger' from 'naeural_core'`.
 - Links: `extensions/serving/default_inference/nlp/th_hf_model_base.py`, `extensions/business/edge_inference_api/text_classifier_inference_api.py`, `extensions/serving/test_th_hf_model_base.py`
+
+- ID: `ML-20260707-001`
+- Timestamp: `2026-07-07T20:44:27Z`
+- Type: `discovery`
+- Summary: EdgeGuard playground stream config can override serving-profile model defaults.
+- Criticality: Operational deployment risk for EdgeGuard model cutovers; source constants and `/model` metadata can report a new target while the active inference stream still loads an older GGUF from persisted stream parameters.
+- Details: During the EGM-029 v0.10 retarget, source defaults and `/model` metadata showed the v0.10 repo/file, but a live generation payload still identified the v0.9 GGUF until the active stream config `STARTUP_AI_ENGINE_PARAMS` was updated. For future cutovers, update both source defaults and the active stream configuration, then verify the returned generation `model` field, not only `/health` or `/model`.
+- Verification: `curl -fsS http://127.0.0.1:5055/model`; generate through the playground server route and inspect the returned attempt `model` field after it calls the model-specific `LLM_INFERENCE_API`.
+- Links: `extensions/business/cybersec/edgeguard/edgeguard_api.py`, `extensions/serving/default_inference/nlp/llama_cpp_edgeguard_qwen_4b.py`
+
+- ID: `ML-20260710-001`
+- Timestamp: `2026-07-10T04:15:31Z`
+- Type: `change`
+- Summary: Moved EdgeGuard cybersec runtime code into a dedicated `extensions/business/cybersec/edgeguard/` package.
+- Criticality: Module-boundary and plugin-discovery change for EdgeGuard API, guard, playground config, and tests.
+- Details: EdgeGuard-specific modules and tests now live outside `red_mesh`; generation is no longer owned by an EdgeGuard LLM-agent plugin or `EDGEGUARD_API /generate`. The playground server route calls model-specific `LLM_INFERENCE_API` workers directly, and `EDGEGUARD_API` stays as the safety facade for model metadata, prompt contract metadata, `/check_cypher`, Neo4j execution, and graph explanation. The serving profile remains under `extensions/serving/default_inference/nlp/` because it is discovered through the AI engine serving-process registry.
+- Verification: `python3 -m unittest extensions.business.cybersec.edgeguard.tests.test_api extensions.business.cybersec.edgeguard.tests.test_cypher_guard extensions.business.cybersec.edgeguard.tests.test_native_api_semaphore_contract extensions.business.cybersec.red_mesh.test_native_api_semaphore_contract extensions.business.edge_inference_api.test_llm_inference_api`; `python3 -m py_compile ...`; `git diff --check`; `importlib.util.find_spec(...)` for the moved EdgeGuard modules.
+- Links: `extensions/business/cybersec/edgeguard/edgeguard_api.py`, `extensions/business/cybersec/edgeguard/edgeguard_cypher_guard.py`, `extensions/business/cybersec/edgeguard/edgeguard_playground.md`, `extensions/serving/default_inference/nlp/llama_cpp_edgeguard_qwen_4b.py`
+
+- ID: `ML-20260716-001`
+- Timestamp: `2026-07-16T14:13:24Z`
+- Type: `correction`
+- Summary: Corrected EdgeGuard graph explanation so Neo4j transport stays in the authenticated Next.js route.
+- Criticality: Security and runtime architecture correction affecting credential scope, Bolt-over-WSS compatibility, explanation availability without the edge-node Neo4j driver, and packet trust boundaries.
+- Details: Corrects `ML-20260710-001` where it implied edge-node owns Neo4j execution for graph explanation. `EDGEGUARD_API` now prepares the validated primary/optional broadening queries and consumes only bounded serialized execution evidence. The Next.js route owns request-scoped credentials and Bolt-over-WSS execution. Edge-node recomputes query/count/flag consistency, rejects connection fields and malformed or oversized graphs, remaps raw graph IDs, sanitizes properties, validates `GraphEvidencePacket`, calls the localhost explanation worker, and validates `CaseExplanation`. Legacy direct-driver mode remains deprecated compatibility behavior only.
+- Verification: `PYTHONDONTWRITEBYTECODE=1 python3 -m unittest extensions.business.cybersec.edgeguard.tests.test_api`; focused EdgeGuard/inference regression suite; `git diff --check`
+- Links: `extensions/business/cybersec/edgeguard/edgeguard_api.py`, `extensions/business/cybersec/edgeguard/tests/test_api.py`, `extensions/business/cybersec/edgeguard/edgeguard_playground.md`, `AGENTS.md`
+
+- ID: `ML-20260729-001`
+- Timestamp: `2026-07-29T21:21:35Z`
+- Type: `change`
+- Summary: Isolated EdgeGuard llama.cpp behavior from generic serving.
+- Criticality: Shared serving-boundary correction affecting generic RedMesh model loading/logging and EdgeGuard runtime identity, determinism, benchmark telemetry, and artifact pinning.
+- Details: Generic `llama_cpp_base.py` is restored to `origin/develop` behavior, including local `MODEL_PATH`, `Llama.from_pretrained`, temperature fallback, retries, and raw output logging. EdgeGuard profiles now inherit a dedicated base that ignores `MODEL_PATH`, requires exact Hugging Face revisions and GGUF SHA-256 values, preserves explicit zero temperature and request seeds, emits the existing fingerprints/benchmark telemetry/context failures, and logs content-free diagnostics. `edgeguard.worker-code-identity.v2` keeps its shape and binds the new shared module through `llama_cpp_base_sha256`.
+- Verification: `python3 -m unittest extensions.serving.test_cybersec_qwen_engine extensions.business.edge_inference_api.test_llm_inference_api extensions.business.edge_inference_api.test_base_inference_api_balancing` (75 passed); `python3 -m py_compile` for changed serving/API tests; `git diff --check`; `git diff --quiet origin/develop -- extensions/serving/default_inference/nlp/llama_cpp_base.py`.
+- Links: `extensions/serving/default_inference/nlp/llama_cpp_edgeguard_base.py`, `extensions/serving/default_inference/nlp/llama_cpp_base.py`, `extensions/serving/base/base_llm_serving.py`, `extensions/serving/ai_engines/stable.py`
+
+- ID: `ML-20260729-002`
+- Timestamp: `2026-07-29T22:45:00Z`
+- Type: `correction`
+- Summary: Removed EdgeGuard-specific llama.cpp serving and returned all three workers to generic serving.
+- Criticality: Corrects the shared serving boundary, local rollout contract, benchmark availability, runtime identity, and output-logging expectations introduced by `ML-20260729-001`.
+- Details: Corrects `ML-20260729-001`: there is no EdgeGuard serving base, adapter, or CyberSec-only engine. The base and finetuned files are configuration-only profiles over the unmodified generic llama.cpp process; CyberSec uses the existing `cybersec_qwen_4b` profile. Local `MODEL_PATH` values select checksum-verified cached bytes operationally, with no runtime revision or SHA enforcement. Health keeps `runtime_fingerprint` and `worker_code_identity` keys but generic workers return `null`. Benchmark mode remains disabled and fails closed at the API gate. Temperature, seed, context-overflow, retry, and generated-output logging follow generic behavior.
+- Verification: `python3 -m unittest extensions.serving.test_cybersec_qwen_engine extensions.business.edge_inference_api.test_llm_inference_api extensions.business.edge_inference_api.test_base_inference_api_balancing extensions.business.cybersec.edgeguard.tests.test_native_api_semaphore_contract` (71 passed); both generic base files match pinned `origin/develop` commit `dc80cab09471f4f64f10b132a43559adcf6dd328`.
+- Links: `extensions/serving/default_inference/nlp/llama_cpp_base.py`, `extensions/serving/base/base_llm_serving.py`, `extensions/serving/default_inference/nlp/llama_cpp_base_qwen_4b.py`, `extensions/serving/default_inference/nlp/llama_cpp_edgeguard_qwen_4b.py`, `extensions/serving/default_inference/nlp/llama_cpp_cybersec_qwen_4b.py`, `extensions/serving/ai_engines/stable.py`
+
+- ID: `ML-20260803-001`
+- Timestamp: `2026-08-03T08:11:49Z`
+- Type: `correction`
+- Summary: Removed inactive EdgeGuard benchmark and attestation scaffolding while preserving an explicit fail-closed API guard.
+- Criticality: Corrects the shared inference API boundary so generic serving is not presented as benchmark-capable or runtime-attested.
+- Details: The historical sealed benchmark remains documentation-only. `LLM_INFERENCE_API` keeps an explicit default-off `benchmark_mode` parameter solely to reject `true` with a stable error and strips `false` before dispatch. Benchmark enablement, seed validation, telemetry handling, health readiness and identity claims, source hashing, and CyberSec worker hashing were removed. Ordinary inference envelopes and the generic llama.cpp serving implementation are unchanged.
+- Verification: Focused inference and serving unit tests; live `edg3` health, disabled-mode rejection, and ordinary completion checks.
+- Links: `extensions/business/edge_inference_api/llm_inference_api.py`, `extensions/business/edge_inference_api/test_llm_inference_api.py`, `extensions/serving/default_inference/nlp/llama_cpp_cybersec_qwen_4b.py`, `AGENTS.md`
+
+- ID: `ML-20260803-002`
+- Timestamp: `2026-08-03T09:53:00Z`
+- Type: `correction`
+- Summary: Removed benchmark mode entirely from the LLM inference API.
+- Criticality: Corrects `ML-20260803-001`; generic inference no longer exposes, validates, rejects, strips, or otherwise interprets a benchmark control.
+- Details: The four LLM completion endpoints have no `benchmark_mode` parameter or special benchmark path. Stale or unknown request input has no supported benchmark semantics. The dormant EdgeGuard UI path remains unavailable because generic workers expose no runtime proof fields. Any reactivation requires a new atomic API/UI/runtime implementation and a new unseen sealed set. The indexed queued-result alignment fix and attributable, content-safe invalid-empty response handling remain as ordinary inference reliability behavior.
+- Verification: Focused balancing, LLM inference, serving-profile, and EdgeGuard API tests; scoped dead-symbol search; exact `llm_utils.py` parity with `origin/develop`; live `edg3` health and ordinary completion checks.
+- Links: `extensions/business/edge_inference_api/llm_inference_api.py`, `extensions/business/edge_inference_api/base_inference_api.py`, `extensions/business/edge_inference_api/test_llm_inference_api.py`, `AGENTS.md`