- Все строки отчёта BENCHMARK_REPORT.md содержат реальный путь, размер файла по os.stat, sha256 первых 64M и имя из general.name
- Отозваны все гипотетические оценки A38; проведены реальные замеры Phi-4 (83.3%, 58.5 ток/с), DeepSeek-Coder-V2 (75.0%, 61.5 ток/с), Granite-4.2-8B (66.7%, 80.6 ток/с)
- Описано падение скорости Multi-head Latent Attention (MLA) DeepSeek на 32k контексте до 3.35 ток/с
- Восстановлены штатные службы владельца на портах 8081 (Qwen3.8-27B) и 8082 (Qwen3-4B)
Preflight and updater tests no longer probe 8081/8082 or GitHub.
Hermetic runs fail-fast on non-loopback sockets. GUI helpers are
importable without customtkinter. CI installs the web extra, and
pytest-timeout plus a wall-clock wrapper bound the suite.
Работа A15 выполнена, но не закоммичена: git в его окружении был
недоступен. Восстановлена ревьюером из рабочего каталога.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`ruff check .` failed on every push since CI was added (1221 findings), so the
job aborted before pytest and the release gate ever ran. Adds [tool.ruff]
selecting bug-catching rules only (E9, F63, F7, F82, F811); style and
modernization rules stay off until their ~1200 findings are burned down.
That selection immediately surfaced seven live defects:
- hermes_hub_app: HubModal used but never imported, so the "Назначить" role
modal raised NameError — the handler added to close an earlier review
finding could never open.
- hermes_hub_app: three `after(0, lambda: ...(e))` callbacks referencing the
except-bound name, which Python unbinds at block exit, so every UI error
path raised NameError instead of reporting the error.
- auto_assigner: build_team_hierarchy referenced an undefined `is_main` and
crashed on every call.
- ui/assets and live_provision_and_validate: annotations naming Any/Tuple
without importing them.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
