feat: harden Responses gateway operations

[clutchpbcfo]

Summary

• remove fabricated zero token usage and mark provider usage as unavailable
• add server request IDs, process-local idempotency replay/conflict handling, and secret-free structured logs
• propagate disconnect cancellation through MCP fetch/polling while failing closed after provider dispatch
• reserve local daily budget durably, commit at create_thread dispatch, and release only before dispatch
• separate liveness/readiness, reject non-loopback binds, and document supported self-host/container operation
• validate official OpenAI JavaScript and Python clients against mocked Hyperagent clients

Validation

• CODEX_BIN="$(command -v codex)" OPENAI_PYTHON_BIN=".../python" npm test — 35/35 passing
• git diff --check
• node --check across source, tests, and JS fixtures
• npm audit --audit-level=high — 0 vulnerabilities
• npm pack --dry-run --json
• targeted secret-pattern scan; only deliberate test placeholders/source field names matched

No Hyperagent credentials, real provider calls, live credits, deployment, or merge were used. Docker is not installed in the validation environment, so the image was not built locally.

Requires Hyperagent platform support

Exactly these remaining guarantees cannot be implemented from the documented OAuth MCP surface:

1. Authoritative token and/or dollar usage returned with thread results.
2. A documented remote thread cancellation operation for client disconnects and timeouts.
3. Provider-backed idempotency plus lookup/reconciliation for ambiguous create_thread outcomes and safe replay across gateway restarts.

Until those exist, the bridge omits usage, cancels local polling only, conservatively commits dispatched budget slots, and fails closed on indeterminate idempotent retries.