Bind Qwen agent tools to the loaded runtime

[andreaborio]

What changed

• bind Agent and Web availability to the model identity confirmed as loaded by the running Hebrus Server
• fail closed on missing, mismatched, switching, stopped, or non-Qwen runtime identity before continuation or network egress
• pin the unified ExpertMajor runtime to Hebrus commit 8015bd39a8d81ebfb997e3955117f481e946a962
• keep Chat and Agents UI aligned with the implemented runtime checks
• preserve the M5 24 GiB AUTO profile and the three-concurrent-Web ordering/citation contract

Why

The previous gate could fall back to the configured model ID even when the loaded runtime identity was missing or different. That allowed Agent or Web policy decisions to be made from intent rather than the runtime actually serving requests.

Validation

• npm audit --audit-level=high — 0 vulnerabilities
• npm run check:brand
• npm run typecheck
• npm test — 46 files, 411 tests
• npm run build
• npm run pack:mac
• npm run verify:mac:dev -- "release/mac-arm64/Hebrus Studio.app"
• post-commit package provenance verified at 4c853e78ac20fd6bebfc3564775d5ee4981f3e91
• git diff --check

Open release gates

• physical M5 24 GiB Sarajevo medium/high, sustained decode beyond 1,719 tokens, and follow-up request
• GLM core gate
• release signing/notarization and final package gates
• source upgrade/rollback E2E still needs explicit authorization for its two npm ci registry requests

This PR must remain draft and must not merge until those gates are closed.