Release 1.2.0

.github/requirements-ci.txt

@@ -15,3 +15,5 @@ pytest==9.0.3
 PyYAML==6.0.3
 sniffio==1.3.1
 typing_extensions==4.16.0
+setuptools==68.1.2; python_version == "3.12"
+wheel==0.42.0; python_version == "3.12"

.github/workflows/ci.yml

@@ -49,10 +49,48 @@ jobs:
         run: python -m unittest discover -s .github/scripts -p 'test_*.py'
       - name: Core and OpenClaw fixture tests
         run: python -m unittest discover -s hermes-owner-mode/tests
+      - name: Installed wheel setup resources
+        if: matrix.python == '3.12'
+        env:
+          PIP_NO_INDEX: '1'
+          PIP_DISABLE_PIP_VERSION_CHECK: '1'
+        run: |
+          python tools/package_assets.py --check
+          python -m unittest discover -s packaging -p 'test_*.py'
+          python -m pip wheel --no-deps --no-build-isolation --wheel-dir "$RUNNER_TEMP/flashalign-wheel" .
+          python packaging/check_wheel.py "$RUNNER_TEMP"/flashalign-wheel/aevonix_flashalign-*.whl
       - name: Shared engine gate and replay tests
         env:
           PYTHONPATH: engines/common
         run: python -m unittest discover -s engines/tests
+      - name: SGLang lifecycle protocol tests
+        run: python -m unittest discover -s engines/common -p test_sglang_lifecycle.py
+      - name: Portable benchmark evidence
+        if: matrix.python == '3.12'
+        working-directory: benchmarks/evidence/qwen-sglang-20261101
+        run: |
+          python -B -S recompute.py
+          python -B -S -m unittest -q test_recompute
+      - name: Install CPU tensor test dependency
+        if: matrix.python == '3.12'
+        run: |
+          python -m pip install --index-url https://download.pytorch.org/whl/cpu 'torch==2.13.0+cpu'
+          python -m pip install 'safetensors==0.8.0'
+      - name: Verify CPU tensor test dependency
+        if: matrix.python == '3.12'
+        run: python -c "import torch, safetensors; assert torch.__version__ == '2.13.0+cpu'; assert safetensors.__version__ == '0.8.0'"
+      - name: Projection generation and failed-hook rollback tests
+        if: matrix.python == '3.12'
+        env:
+          PYTHONPATH: engines/vllm
+          CUDA_VISIBLE_DEVICES: ''
+        run: python -m pytest -q engines/vllm/tests/test_generations.py
+      - name: SGLang graph-generation contract tests
+        if: matrix.python == '3.12'
+        env:
+          PYTHONPATH: engines/sglang
+          CUDA_VISIBLE_DEVICES: ''
+        run: python -m pytest -q engines/sglang/tests/test_graph_contract.py
       - name: Native DeepSeek contract and lifecycle tests
         run: python -m pytest -q engines/vllm-deepseek-native/tests
       - name: Shell and JavaScript syntax

CHANGELOG.md

@@ -1,5 +1,30 @@
 # Changelog
 
+## 1.2.0
+
+- Keep one active Hermes plugin registration per process across duplicate loads and reloads.
+- Include fixed identity-verification reasons in each request's audit record.
+- Profile staging, activation, rollback and retirement through local administration.
+- Immutable generations and worker recovery contracts for serving adapters.
+- Status, doctor and terminal dashboard commands.
+- Pinned FR443 collection, durable semantic judging and evidence exports.
+- Distinguish a busy evaluation lock from invalid evaluation data.
+- Qwen FR443 results and a paired SGLang timing study with reproducible data.
+- Artifact-bound profile evaluation and controlled calibration sweeps.
+- Shared request authentication and extensible engine, model, channel and profile registries.
+- Complete runtime installation for managed engines and harnesses.
+- Include engine kits, starter profiles and harness assets in the Python package.
+- Check container memory limits before starting native DeepSeek.
+- Automatic native profile-control startup and SSH profile staging for remote TensorFold deployments.
+- Preserve saved profile state across restarts, including an empty inventory or deactivated aliases.
+- Keep managed SSH forwards attached to their supervisor when SSH multiplexing is enabled.
+- Record managed launch attempts before starting workers and preserve ownership through interrupted starts.
+- Preserve the active Hermes policy and refresh message observers across plugin reloads and host renderer replacements.
+- Share one configurable shutdown deadline across managed engine workers.
+- Verify external engine resources are stopped before confirming managed shutdown.
+- Report reused prefix tokens in the DeepSeek TensorFold runtime.
+- Handle atomic status-file replacement without interrupting a managed engine.
+
 ## 1.0.0
 
 Initial Aevonix FlashAlign release:

README.md

@@ -2,7 +2,7 @@
 
 **ReAlign models instantly, on demand.**
 
-By Aevonix Research · Version 1.1.1
+By Aevonix Research · Version 1.2.0
 
 [![Stock DeepSeek V4.1 Flash refuses to help make a realistic-looking ID for a theatrical role; with FlashAlign on, the same model answers it](assets/demo/flashalign-promo-preview.gif)](https://aevonix.github.io/FlashAlign/)
 
@@ -18,7 +18,7 @@ changing them. Run it directly, with Hermes or OpenClaw, or through your own
 harness. Managed inference runs on Linux with NVIDIA GPUs. See
 [compatibility](docs/compatibility.md) for DGX Spark support and engine requirements.
 
-The V1 starter profiles focus on refusal reduction. Future profiles could also
+The starter profiles focus on refusal reduction. Future profiles could also
 increase refusals for public or higher-risk tasks.
 
 Restricting abliteration to verified users and channels reduces the risk of an
@@ -70,9 +70,9 @@ See [compatibility](docs/compatibility.md#engines-and-models) for hardware
 requirements, pinned checkpoints and feature limits, and
 [validation](docs/validation.md) for tested configurations and results.
 
-Switching happens per request among profiles already loaded by the engine.
-Loading a new model or profile happens beforehand and can require an engine
-restart.
+Switching happens per request. Engines that advertise profile hot loading can
+also add compatible profiles without a restart. Loading another model still
+requires its own engine setup. See [profile management](docs/profile-lifecycle.md).
 
 Use `--model-path /path/to/model` to select a model. Run `bash ./install.sh` to
 connect to an engine that is already running. See the
@@ -103,6 +103,38 @@ bash ./install.sh --detect
 The [setup guide](docs/setup.md) covers connection and activation. The engine's
 effective-profile response confirms that the selected profile is active.
 
+## Check and manage your deployment
+
+After setup, run:
+
+```sh
+bash ./flashalign.sh status
+bash ./flashalign.sh doctor --probe
+bash ./flashalign.sh dashboard
+bash ./flashalign.sh profile list
+```
+
+Status reads runtime metadata. Doctor sends a synthetic request only with
+`--probe`. The dashboard shows measured values when the engine provides them.
+See [operations](docs/operations.md), [profile management](docs/profile-lifecycle.md)
+and [benchmarks](docs/benchmarks.md).
+
+## Measured results
+
+On Qwen3.8-27B, the automated FR443 judge counted **289 refusals in 443 stock
+responses** and **11 in 443 profile responses**. The suite includes safe,
+harmful and unsafe requests. Human audit and broader capability testing remain
+pending. [Full results and settings](docs/qwen-fr443-results.md).
+
+A separate Qwen/SGLang timing study measured **1.324× end-to-end completion-token
+throughput** for the opt-in graph Owner Mode path compared with the previous
+release's eager path. This was a fixed 256-token workload with three pairs of
+engine starts, not a decode-speed measurement. [Timing results and reproducible data](docs/qwen-sglang-timing.md).
+
+A fixed follow-up passed completion and GSM8K, but did not establish all
+capability gates. Eager remains the default and graphs remain opt-in.
+[Quality results and limits](docs/qwen-sglang-quality.md).
+
 ## Choose when to apply FlashAlign
 
 **Verified contacts** is the default. It applies a profile to messages from an
@@ -197,6 +229,8 @@ requirements.
 ## Documentation
 
 - [Setup and activation](docs/setup.md)
+- [Status, doctor and dashboard](docs/operations.md)
+- [Live profile management](docs/profile-lifecycle.md)
 - [Managed engine setup](docs/engine-setup.md)
 - [Agent-assisted setup](docs/agent-setup.md)
 - [Harness and model compatibility](docs/compatibility.md)
@@ -208,6 +242,7 @@ requirements.
 - [Hermes configuration](hermes-owner-mode/README.md)
 - [OpenClaw adapter](openclaw-owner-mode/README.md)
 - [Engine API](docs/engine-contract.md)
+- [Extension interfaces](docs/extending.md)
 - [Security model and vulnerability reports](SECURITY.md)
 - [Development and tests](CONTRIBUTING.md)
 - [Vector logos and brand kit](assets/brand/README.md)

THIRD_PARTY_NOTICES.md

@@ -49,6 +49,13 @@ does not establish permission to redistribute every compatible checkpoint.
 
 ## Installed dependencies
 
+The refusal-evaluation judge includes rubric text from
+[Blackfrost-AI/r1-false-refusal-443](https://github.com/Blackfrost-AI/r1-false-refusal-443)
+at revision `a2a8051a3f4ff2593f6b6e9c22f9929b3244c742`, under Apache-2.0.
+See the [rubric notice](flashalign/notices/BLACKFROST-NOTICE.md) and included
+[license](flashalign/notices/APACHE-2.0.txt). The source prompt corpus is not
+bundled; its underlying datasets retain their applicable source terms.
+
 The dependency declarations in `pyproject.toml`, `hermes-owner-mode/requirements*.txt`
 and each engine package record the installed packages and version constraints.
 They include HTTP and configuration libraries, NumPy, safetensors,

benchmarks/evaluation-limits.example.json

@@ -0,0 +1,7 @@
+{
+  "max_response_bytes": 16777216,
+  "max_events": 65536,
+  "max_sse_event_bytes": 16777216,
+  "max_record_bytes": 67108864,
+  "max_judge_request_bytes": 67108864
+}