perf(runtime): overlap connection preflight and trace preparation stages

[sam2tom]

Runtime startup currently waits for Harness discovery before starting authenticated bootstrap, even though those operations are independent after credentials and enrollment are resolved. Run them concurrently, require both to succeed before registration, and cancel and join the sibling on failure.

Add trace-correlated preparation timings for workspace setup, model catalog/session planning, process spawn, RPC initialization and verification. Preserve model validation, pinned catalogs, process ownership and reconnect behavior. Logs omit native error text, command output and credentials; the operator guide explains nested timing spans.

Validation:

• Affected CLI, Codex and dispatch package tests and race checks passed; startup race checks repeated ten times.
• Runtime contract target, go vet, three translation tests, Linux amd64 CLI cross-compilation and diff checks passed.
• Full local hygiene and website build/test gates passed, including 22 website tests.
• Two fresh independent reviews, including the final CI corrections: zero actionable findings.
• A separate patch against deployed beta was validated, including its one-shot pairing order and selected-Harness behavior; this PR targets main only.
• Synthetic two-by-5ms wait benchmark: serial approximately 11.1ms, concurrent approximately 5.6ms. This measures dependency overlap only, not production latency.

Also remove the obsolete Codex builder override so the template workflow uses the existing build-codex-runtime.sh path and passes the name guard. No SQL, dependency pin or deployment changes. Activation requires rebuilding and qualifying the Runtime template; deploying Core alone does not update Runtime binaries or existing sandboxes. Real-provider acceptance and the cold-first-text <=12s target remain unverified. No template build, activation or deployment is included.

Related cross-repository work for joint review: https://github.com/sandbaseai/sandbase-monorepo/pull/1067. Neither PR establishes live cold-start acceptance.

Final CI: core-check passed all 25 jobs at 81bd4d7, including backend/integration, official-client, native platforms, Compose, website and Web acceptance. Optional credential-dependent live tests are not production acceptance.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.