BUG: Event-loop blocking -- sync CPU/IO in async paths not caught by sync_in_async

[echel0nn]

Labels: bug, async, performance
Severity: S1. The sync_in_async honesty rule only detects session_scope(); these blocking calls pass the audit.

Finding (path:line)	Issue
platform/tools/knowledge.py:77,169	KnowledgeStoreTool.forward / KnowledgeQueryTool.forward are async def and call model.encode(...) (synchronous CPU, BGE-M3/MiniLM, 100ms-2s on CPU) inline on the event loop; _get_embedding_model() also loads the model synchronously on first call. Since the forward is async it cannot be to_thread-wrapped by the dispatcher
platform/services/embedding.py:50,83	BGEProvider.embed / MiniLMProvider.embed are sync model.encode(...); KnowledgeService calls them from async paths
storage/database.py:230 (backup_database)	async def runs subprocess.run(["pg_dump", ...], timeout=300) synchronously -> blocks the loop for the full dump (up to 300s)
modules/forensics/services/offline_installer.py:197,208,327	async def prepare_pip_wheels / _install_apt_offline run subprocess.run(["pip"/"apt-get", "download", ...], timeout=600/120) synchronously on the loop
modules/vulnerability/providers/nvd.py:95,144 (+ epss.py, kev.py)	Sync httpx.Client + time.sleep() in the async provider chain (reinforces #55)

Architectural inconsistency (module to_thread ban): HONESTY_AUDIT.md bans modules under src/aila/modules/ from asyncio.to_thread/ThreadPoolExecutor, so modules/vulnerability/reporting/pdf.py:104 and modules/vr/reporting/pdf_report.py:186 run weasyprint (CPU-bound, no async API) on the event loop by policy. But modules/malware/services/target_analysis.py:1161, modules/vr/services/target_analysis.py:1148, and modules/vr/api_router.py:3527 (MASVS PDF) all call asyncio.to_thread -- either whitelisted or the audit rule has a gap. The rule is inconsistently enforced, and CPU-bound module work has no consistent offload story.

Acceptance: move model.encode, pg_dump, and pip/apt download off the loop (sync helper + platform-boundary to_thread, or an async driver); provider I/O uses httpx.AsyncClient + asyncio.sleep; broaden the sync_in_async audit rule to also flag subprocess.run, time.sleep, sync httpx.Client, and model.encode inside an async def body; settle the module to_thread policy (either allow a sanctioned offload helper or enforce the ban consistently).

[echel0nn]

Design decision (2026-07-19). The broadened sync_in_async honesty-audit rule is APPROVED for CI: flag subprocess.run, time.sleep, sync httpx.Client, and model.encode inside an async def body, not just session_scope().

[echel0nn]

Implementation-ready design (2026-07-19). Wave 1b. Full design on disk: .run/designs/DESIGN_async_correctness.md (8-heading template: grounded problem, design with signatures+DDL+Alembic, reworked concerns, test plan, rollout).

Key moves:

• Broaden sync_in_async to flag subprocess.run/time.sleep/sync httpx.Client/model.encode in async; escape via run_blocking_io.
• Wrap model.encode, pg_dump, pip/apt in to_thread; providers move to async I/O.