Optimize auth/request hot paths and codex instruction flow

[iam-brain]

Summary

• Reduce hot-path overhead by making quota refresh non-blocking, consolidating outbound payload transforms, and skipping no-op auth/snapshot writes.
• Tighten codex instruction behavior: plan prompt applies only to plan profile, build replacement is codex-only and independent of collaboration profile, and orchestrator/plan tool calls are replaced without runtime compatibility block injection.
• Remove unused collaboration tool-profile config surface and update schema/docs/tests for the simplified behavior.

Verification

• npm test -- test/config.test.ts test/codex-native-spoof-mode.test.ts test/codex-native-in-vivo-instructions.test.ts test/codex-native-request-transform.test.ts test/openai-loader-fetch.prompt-cache-key.test.ts test/codex-status-storage.test.ts test/acquire-auth-locking.test.ts test/codex-prompts-cache.test.ts test/orchestrator-agent.test.ts
• npm run typecheck
• npm run build
• npm run verify