feat(runtime): add native PHP backend

[chubes4]

Closes #2473

Summary

• add an opt-in wordpress-native backend with a built-in contained Docker driver, alongside optional backendPackage driver injection
• keep recipes backend-neutral: the same wordpress.run-php, wordpress.browser-actions, and wordpress.bench commands select Playground or native execution
• run a digest-pinned linux/amd64 WordPress/PHP 8.4 Apache container with persistent OPcache and two prefork workers, plus a disposable tmpfs MariaDB sidecar
• keep the database on an internal Docker network while the app publishes on loopback only, so the database is never externally reachable
• install WordPress through its own API with generated fixture-only credentials, then execute validated steps-json browser interactions with a real form login
• emit machine-readable provenance and benchmark evidence explicitly marked as local representative measurements, not production RUM
• deterministic teardown removes both containers and both networks, including after partial startup

Live verification on a Docker runner

npm run test:native-docker-runtime passed against real containers (runs wp-codebox-2473-native-docker-live-3e41c332 and wp-codebox-2473-native-live-evidence-563f7d52).

Captured evidence:

• browser steps: navigate and expect #wpadminbar both passed after an authenticated wp-login.php submission
• network: WordPress 7.1 admin assets served over the published loopback port
• page errors: none
• benchmark: coldStartupMs: 3785, warmNoopPhpMs: 54, dynamicWordPressRequestMs: 32

For reference, #2473 measured Playground at 34.8 s cold and 0.5–0.7 s warm for stock WordPress, so this backend provides the intended performance lane.

Other verification

• npm run build
• npm run test:native-runtime
• npm run test:runtime-services
• npm run check
• git diff --check

Known limitations

• Live containment requires Docker; without it the integration lane skips explicitly as trusted-containment-tools-unavailable rather than falling back to host PHP.
• Readonly directory mounts are one-time source snapshots copied into the container, not live Docker bind mounts; unsupported mount modes fail explicitly.
• A matched Playground-vs-native benchmark table still needs to be collected in one Docker-capable environment; only the native side is measured here.
• npm run check also exercises tests/agent-runtime-components.test.ts, which is sensitive to locally discovered components outside this change.

AI assistance: OpenCode with openai/gpt-5.6-terra and openai/gpt-5.6-sol was used to inspect the runtime architecture, implement the backend, and run live containerized verification.