[codex] keep Hermes task-plan MCP on its owning Studio instance

[EKKOLearnAI]

Summary

When Desktop and another Studio server share a Hermes profile, managed MCP auto-injection can overwrite the URL in that profile. A subsequently started Desktop worker then submits its valid task-plan context to the other server and receives HTTP 409 (Task plan context is unavailable or has expired).

Bind managed MCP subprocess launches to the Bridge owner's Studio URL/state directory and the worker's profile. Apply the override to a copy at launch/reconnect, preserving shared YAML and user-owned MCP settings. Cover split and legacy Hermes MCP layouts, and extend model-switch coverage to verify that another running session stays unchanged.

Existing Bridge/MCP workers need a Studio restart after installing the fix. This does not restore previously failed turns.

Validation

• 194 focused Vitest tests passed across MCP instance routing, Bridge manager/imports/readiness/final context, Python concurrency, provider credentials and task-plan lifecycle.
• Real MCP subprocess test reproduces 409 against the wrong HTTP server and succeeds against the owner with the same stale launch configuration.
• Actual Hermes 0.20.6 and 0.21.3 MCP environment builders verified locally with isolated temporary homes.
• PLAYWRIGHT_PORT=14173 npx playwright test tests/e2e/task-plan.spec.ts --workers=2 — 4 passed.
• npm run harness:check — passed.
• npm run build — passed.
• git diff --check — passed.