refactor(aiops): unify versioned runtime and dynamic capabilities

[singchia]

Summary

Unify Agent execution around versioned Eino Graph runtimes while retaining chat, delegated workers, approvals, workflows, MCP, skills and built-in tools.

• Isolate v1 (first release v0.7.158), v2 (v0.18.2) and v3 sources. Wire selectable v2/v3 runtimes and mark v1 unsupported. New sessions follow the administrator setting; existing sessions, child workers and approvals retain their runtime owner. Legacy empty versions stay on v3.
• Remove implicit 15-second execution limits while preserving cancellation, connection probes, approval expiry and explicit dispatch budgets.
• Add per-turn skill catalogs and v3 load_skill, retain always/keyword activation, and support on-demand skills. Pin package resources for active turns/approvals; uninstall revokes new acquisitions. v2 uses compatible skill-body injection, not v3 resource leases.
• Refresh dynamic MCP catalogs through immediate registration wakeups, paginated discovery, change notifications and periodic polling. Reject stale publication, revoke disabled/removed tools, and bind approvals to registration, credentials and schema revisions.
• Preserve built-in agents across registry reloads. Reconcile interrupted internal workers on single-Manager startup with transactional tool finalization and interruption evidence, without replaying external work or changing independent approval records.
• Distinguish search success, partial results, empty results and failures. Validate provider settings and use only explicitly configured fallback providers.

Fixes #498. Refs #499. Design and evidence: RFC-008, RFC-009, PRD-010, docs/test/e2e-catalog.md and docs/test/runtime-v2v3-deploy-20261017-5.md.

Validation

• Full isolated backend E2E with race: 31 top-level / 95 total tests passed, zero failures/skips. Real Manager/MySQL with simulated model/MCP services.
• Affected runtime, registry, MCP, settings, approval and startup/storage race suites passed. Latest worker recovery tests cover scope, ownership, idempotence and transactional rollback; targeted go vet passed.
• Chrome E2E: 21 passed, 1 skipped, zero failures. Kubernetes live resource validation is skipped because the test environment has no cluster. Kubernetes code acceptance passed 51 frontend tests and all four Go layers with race.
• Test environment Manager .8 / frontend .5: six live-model scenarios passed, covering v2/v3 delegated cancellation, v2 skill fallback, legacy empty-version ownership and v2/v3 workflow worker interruption on restart. Workers close, pending tools finalize, interruption records remain unique, and external replay count is zero. Temporary fixtures are cleaned and the default restored to release/v3.
• Earlier live validation also passed dynamic MCP discovery/calls, skill install/load/uninstall, preserved default-agent configuration, SSE disconnect/history continuation and v2 MCP approval execution across a Manager update.
• Linux runtime build/race, frontend build, settings UI checks and git diff --check passed. Latest GitHub checks are authoritative for this PR head.
• Earlier broad container testing recorded three failures in unchanged paths involving AutoAPM worker resources, host worktree Git metadata and inaccessible-file permissions. Global frontend lint previously reported 9 errors/43 warnings outside the changed paths. These are disclosed separately from the passing targeted and E2E suites.

Risk and rollback

Merge is authorized after verification; production deployment remains separate. The test deployment is at https://101.34.63.91/.

Runtime selection changes only new sessions. Keep both v2/v3 engines available for existing sessions and approvals; switching the default is not an old-image rollback. v1 is unsupported. Skill resource leases remain process-local, so stale approvals fail explicitly after restart. Interrupted workers terminate with evidence rather than automatically resuming. Startup reconciliation assumes a single Manager; multi-Manager ownership/leases are not implemented. Chat AgentTool does not expose background=true; background acceptance uses workflow Agent nodes.

Dynamic capability and permission drift remain the principal risks, covered by lifecycle/concurrency tests. Search fallback is opt-in. Current test rollback: python3 /opt/ongrid/test-deploy-20261017-worker-recovery-8/ongrid-deploy8.py rollback restores Manager .7, preserving recorded interruption state. No schema change in the latest deployment; retain compatible runtime ownership columns and never overwrite newer data for an application rollback.

Author confirmation

• I have read and agree to the project contribution guide. Guide: https://github.com/ongridio/ongrid/blob/main/CONTRIBUTING.md

[github-actions]

Commit message policy

Please write the PR title and every commit message in English using Conventional Commits, for example fix: handle empty service names. The automated check validates the format and rejects Han characters.

See the latest PR checks and open commit-policy for the current result and details.

This reminder is posted once per PR. Later failures and fixes only update the checks; this comment is not updated.