Self-hosted server crashes silently during second agent deploy via CLI

## Bug Report

### Description

The self-hosted Blink server (`blink-server` Docker container) crashes silently and repeatedly when deploying a second agent via the CLI. The first agent was deployed successfully through the UI. Subsequent deploys via `blink deploy` (from a GitLab CI/CD pipeline) cause the server process to die with no error output, no stack trace, and no OOM indication.

### Environment

- **Blink Server Version:** `0.1.4+3aa5336`
- **Deployment:** Self-hosted, Docker (`ghcr.io/coder/blink-server:latest`)
- **Database:** Amazon RDS (Postgres, external)
- **Host:** EC2 instance, 3.7GB RAM
- **Node.js heap limit:** ~1965 MB
- **Docker memory limit:** None (`0`)
- **Agent bundle size:** ~14MB (`agent.js` is 13,978,840 bytes)

### Steps to Reproduce

1. Deploy Blink server via Docker with Postgres on RDS
2. Successfully deploy a first agent via the web UI (this works fine)
3. Create a second agent and deploy via CLI (`blink deploy` from CI/CD pipeline)
4. Server crashes silently during the deploy

### Expected Behavior

The server should deploy the second agent container and report success, or fail with an error message.

### Actual Behavior

The server logs show:
```
Deploying agent ()
```

Then immediately crashes with no further output. No error, no stack trace, no "Writing files to" log line (which should be the next log statement in `deployAgentWithDocker`).

Docker restarts the container (via restart policy), the server picks up the pending deployment from the DB, and crashes again — creating a crash loop. Each CLI retry also creates a new agent record, resulting in orphaned agents.

### Diagnostics Performed

- `docker inspect blink-server --format '{{.State.OOMKilled}}'` → `false`
- `docker inspect blink-server --format '{{.HostConfig.Memory}}'` → `0` (no limit)
- `docker exec blink-server node -e "console.log(v8.getHeapStatistics().heap_size_limit / 1024 / 1024 + ' MB')"` → `1965.25 MB`
- Host has ~2.9GB available RAM, existing containers using ~315MB combined
- Only two containers running: `blink-server` (~180MB) and one existing `blink-agent` (~135MB)

### Analysis

The crash occurs between the `"Deploying agent"` console.log in `packages/server/src/agent-deployment.ts` and the `"Writing files to"` log. This narrows it to either:
- `querier.updateAgentDeployment()` (status → "deploying")
- `downloadFile()` → `db.selectFileByID()` (reading the ~14MB file from RDS as a bytea column)

The silent crash with zero error output suggests either an unhandled promise rejection that kills the process before the error handler fires, or a crash in the Postgres client when handling a large binary column.

### Workaround

None identified yet. The first agent deployed via UI works fine, suggesting the issue may be specific to the CLI deploy path or related to deploying a second agent alongside an existing one.

### Additional Context

The CLI `blink deploy` successfully:
- Builds the agent (esbuild, 2107ms)
- Uploads all 11 files (13.3MB total) to the server

The server receives the deploy request and begins processing, but crashes during the file download/write phase.