Release 2.2.6: server load and shared-service hardening

[gaochangw]

Studio users can now open Server load to see active-session estimates, running/queued jobs and separate API/compute CPU, memory and GPU panels. One persistent UTC cleanup timestamp replaces the repeated page alerts, including a compact mobile layout.

The public service serializes workspace mutations, caps expensive analysis/generation work, reserves temporary storage and keeps cleanup alive after GPU-finalization failures. Local/private boundaries reject ambiguous headers and malformed origins, sessions expire server-side, CSRF comparisons are constant-time, and authenticated telemetry contains only aggregates. This builds on the filesystem/disclosure fixes merged in #38.

Runtime dependency floors and the pinned PyTorch 2.14/CUDA 13.2 image are updated. The image receives available Ubuntu fixes and drops unused build/install tools; CI audits frontend and resolved Python dependencies. README, guides, current screenshots and a dated security report explain scope and remaining findings.

Validation:

• Complete local Python run: 951 passed; the single API snapshot comparison passed after the new nullable fields were exported. Final boundary/GPU/public/API regression: 79 passed, plus the added duplicate-header/non-ASCII CSRF case.
• Frontend: 234 tests, TypeScript, lint, production build and generated contract checks passed. Strict docs and wheel/sdist metadata checks passed.
• Actual Chromium at desktop and 390/360 px: one cleanup timestamp, no overflow/collisions, stale readings hidden, polling stops after navigation.
• Actual isolated CUDA container: PA, DPD and ILC/ILA succeeded on 32,768 synthetic pairs with 6,452 separate test samples.
• 100 cached loopback status requests: median 1.01 ms, P95 1.56 ms. This is local endpoint timing, not an Internet/capacity benchmark.
• Frontend/installed Python advisory checks clean. Container scan: 0 high/critical, 36 remaining medium/low OS findings without vendor fixes; full unsuppressed summary is committed.

Release: 2.2.6. After checks and review, deploy the same source to API, GPU agent/image and Pages with a drained queue and rollback copies, then verify the public workflow and published wheel. No claim of independent penetration-test certification, MPS hardware validation or physical RF performance.

[gaochangw]

Self-review of f1a4227 completed. I traced authentication/admission through the outer public boundary and nested local app, checked counter/storage release on errors, verified that GPU cleanup failures leave admission/health closed without stopping tenant expiry, and checked telemetry privacy, stale values and UI polling. The updated image contains the same 190 Python source files as this PR and passed real isolated CUDA PA/DPD/ILC runs.

The desktop/mobile views, package startup, contract, docs, security regressions and dependency scans pass locally. All three default-install CI jobs, package build, docs and guard already pass; the full Python/browser jobs are still running and must pass before merge. No remaining blocking code findings. The dated report explicitly retains 36 unfixed medium/low Ubuntu findings and the host/driver/shared-GPU limits; it does not claim a complete independent penetration test.