refactor: flatten runner and vertical access, remove 1029 lines

[lbx154]

The CLI adapter loaded its only supported runner through a string-keyed dependency dictionary and probed fields to accommodate obsolete test doubles. Vertical consumers repeatedly converted the same provider through individual accessor functions. Use the bundled runner's real types and read the validated VerticalContract directly, reusing it within each operation.

• Remove the runtime loader and redundant accounting/I/O delegates; share consecutive tool-call detection across Claude and Cursor events.
• Move automatic stage-completion validation onto VerticalContract. Keep the legacy accessor names imported by the current community argus-verticals repository; remove the remaining unused forwarding API.
• Replace the fake runner package with real types and an execution-boundary mock. Consolidate repeated metering, watchdog, reconnect, startup-refusal, and provider-cost cases without dropping their input scenarios.
• Net reduction from dev 450b3138: 1,029 lines (1,530 removed, 501 added), including 376 production Python lines. No code relocation or generated-file deletion contributes to that total.

Validation: Ruff and git diff --check pass; npm run check passes all 226 tests. The local full Python suite completed with 11,383 passed and 97 skipped. Four install-state tests conflicted with the shared environment’s preinstalled community package; all four passed in a fresh virtual environment. CI runs the full suite without that extra package. All 24 usage-record fixture scenarios were compared against the baseline to verify identical effective inputs.

Documentation: update the runtime maintenance guide with the direct loader/contract paths and compatibility scope. CI/workflow: no workflow changes; require the existing Linux Python suite and Linux/macOS/Windows TypeScript and ownership checks before merging. No service deployment or restart is included.