# G5 cutover slice live — first real Bun-frontend traffic to Go substrate

2026-05-01: the Bun mcp-server (which serves devop.live/lakehouse/
via nginx → :3700) now has a `/_go/*` pass-through that routes
requests to the Go gateway. Real frontend traffic can flow through
the Go substrate while the existing /api/* path (Rust gateway)
stays untouched.

## What changed

**Rust legacy repo** (`/home/profit/lakehouse/`):

- `mcp-server/index.ts`:
  - New `GO_BASE = process.env.GO_LAKEHOUSE_URL || ""` (off-by-default)
  - New `/_go/*` handler at the same shape as `/api/*` (forwarded
    method+headers+body) but pointing at `${GO_BASE}${path}`
  - Returns 503 with rationale when `GO_LAKEHOUSE_URL` is unset
  - Returns 502 when Go gateway is unreachable (preserves operator
    visibility into Go-side failures vs silent fallback)

**System config**:

- `/etc/systemd/system/lakehouse-agent.service.d/go-cutover.conf`:
  ```
  [Service]
  Environment=GO_LAKEHOUSE_URL=http://127.0.0.1:4110
  ```
  Drop-in adds the env var to the systemd-managed Bun service.
  Reversible via `systemctl revert lakehouse-agent.service` or by
  removing the file.

**Go rewrite repo** (this commit): documentation + cross-runtime
parity log entry.

## Verification (all live, against persistent Go stack on :4110)

| Path | Routes to | Verdict |
|---|---|---|
| `GET /_go/health` | Go gateway `:4110/health` | `{"status":"ok","service":"gateway"}` |
| `POST /_go/v1/embed` | Go embedd via gateway | nomic-embed-text-v2-moe vectors, dim=768 |
| `POST /_go/v1/matrix/search` | Go matrixd via gateway | 3/3 Forklift Operators against the 200-worker persistent corpus |
| `GET /api/health` (control) | Rust gateway `:3100/health` | `lakehouse ok` (unchanged) |

Search results from the live cutover slice:

```
rank=0 id=w-43  dist=0.445 Brian Ramirez   (Forklift Operator, Springfield, IL)
rank=1 id=w-102 dist=0.448 Laura Long      (Forklift Operator, Cleveland,   OH)
rank=2 id=w-101 dist=0.485 Terrence Gray   (Forklift Operator, Champaign,   IL)
```

3/3 role match. Top-1 in IL exactly. Real coordinator-shape query
served by Go substrate via Bun frontend.

## What this is

This is the **first time** real Bun-frontend traffic has reached the
Go substrate end-to-end. Production paths (devop.live/lakehouse/* →
nginx → Bun :3700 → /api/*) still go through Rust on :3100; the
new /_go/* path is a parallel slice that operators or external
tools can hit to validate Go under realistic frontend conditions.

The cutover unit is FULLY ADDITIVE:
- No existing tool modified
- /api/* unchanged
- All Bun mcp-server tools that currently call BASE (Rust) unchanged
- /_go/* fails-closed (503) when env var unset

The cutover unit is FULLY REVERSIBLE:
- Unset `GO_LAKEHOUSE_URL` → /_go/* returns 503
- `systemctl revert lakehouse-agent` → removes the drop-in
- `git revert` of the Bun source change → removes the handler

## What this is NOT

- **Not** an nginx flip. devop.live/lakehouse/* still goes
  through Rust. Operators have to opt into /_go/*.
- **Not** a wholesale routing change. Each path under /_go/ is a
  manual choice; no automatic routing of Rust paths to Go
  equivalents.
- **Not** a transformation layer. /_go/v1/embed expects the Go
  request shape (`{texts, model}`); /_go/* doesn't translate
  Rust-shaped requests for the caller.

## Operational discipline

The systemd drop-in is at `/etc/systemd/system/lakehouse-agent.service.d/`.
Operators inspecting the system should know to look there for
cutover-related env overrides.

The persistent Go stack must be up at :4110 for /_go/* to succeed.
Tear down via `pkill -f 'bin/persistent-'`. Bring up via
`scripts/cutover/start_go_stack.sh`.

## What's next

- **Pick a real coordinator-facing tool to flip.** The current
  /_go/* pass-through is operator-facing (manual curl). The next
  step would be modifying ONE specific tool in mcp-server/index.ts
  (e.g. `/search` → /v1/matrix/search) to optionally route through
  Go when GO_BASE is set, with response shape adapter. That's a
  product-visible flip; current state is infrastructure-visible.
- **Shadow-read mode.** A hybrid where /api/* ALSO calls /_go/*
  in parallel and logs diffs to stderr — operators see Go-vs-Rust
  divergences on real traffic without changing user-facing
  behavior. Cleaner cutover discipline than direct flip.
- **Production load test.** This slice has been exercised on 3
  manual curl requests. A traffic generator hitting /_go/v1/matrix/search
  at sustained rate would expose latency / stability under load
  that single-shot tests can't.