fix(studio): preserve domain Skills and MCP across runtime updates

[zyn080302]

Updating Skills or MCP servers in Studio can replace a domain application's tool wrappers, lose previously selected Skills, or fall back to regenerating application source. This change completes the source-preserving update path so existing domain runtimes retain their entry point, locked dependencies, dynamic prompts, tool policies, identity, and protected configuration across repeated edits.

The update uses a digest-pinned base image plus a small Skill/MCP overlay. It respects the application's MCP compatibility handler, validates public overlay metadata, preserves selected Skills across successive builds, and keeps credential values out of images and browser storage. Existing source-preserving runtimes do not require a new veadk.aiops dependency or migration to Managed shared. Normal creation and regenerated-source updates retain their existing behavior. Frontend assets and mandatory regression gates are included.

Validation recorded on 2026-10-14 against base a7569a04 and commit 5f5f647f:

• Mechanism regressions reproduced the failures before the fix; 339 affected Python tests passed after synchronization. Three offline tests using the actual database tool factory verified Skill execution, MCP selection, and rejection of SQL writes before backend access.
• Frontend: 1,386 Node tests, 72 Vitest tests, and the production build passed. Backend release gates passed: 149 coverage tests / 1 skip at 96%, 9 lifecycle tests, 233 credential tests, and 39 release tests. Ruff and secret scans passed.
• Two real Studio edit/publish cycles on the authorized legacy V4 database runtime completed successfully (V18 and V19). Reopening the editor preserved Skills, MCP settings, and configured credential references. Both real requests produced complete reports and exercised the selected tools.
• Validation limits: existing Pyright diagnostics remain (38 on the current tested dependency set; no additional diagnostics). The full Python run reported 7,063 passed, 53 skipped, 4 expected failures, and 3 wheel-fixture setup errors; the three failed setups passed on an unchanged isolated rerun. This was not a clean first full-suite run.
• Customer V4-2 runtime updates have not been exercised. This PR does not publish a Studio upgrade channel or deploy any customer runtime. Upstream main has advanced since the recorded test base; CI must evaluate the current integration result.

Design and detailed evidence: English / Chinese.