fix. harden Go review flow

Fix runner bootstrap and auth handling, preserve queued SHAs, make event/job acceptance atomic, fence stale runs, correct retries and prompts, add fake end-to-end coverage, and fix deployment defaults.

Co-Authored-By: Claude <noreply@anthropic.com>

README.md

@@ -88,7 +88,7 @@ cp .env.example .env
 docker compose -f docker-compose.dev.yml up --build
 ```
 
-The bot container needs access to the host Docker API to launch isolated review containers. Mounting `/var/run/docker.sock` is a privileged deployment decision; use a dedicated runner service or hardened Docker host where possible.
+The bot container needs access to the host Docker API to launch isolated review containers. Mounting `/var/run/docker.sock` is a privileged deployment decision; use a dedicated runner service or hardened Docker host where possible. `docker-compose.yml` consumes the published image, while `docker-compose.dev.yml` builds locally. ChatGPT auth mode should use a private Compose override to mount `auth.json`; the checked-in Compose files do not mount it in API-key mode.
 
 ## Repository configuration
 
@@ -108,7 +108,7 @@ ignore:
   - generated/
 ```
 
-The file is read from the PR head and is treated as untrusted data. It cannot choose commands, credentials, images, host paths, container privileges, or network policy. Tests are disabled by default; `tests` mode or `include_tests: true` explicitly permits the runner to execute project tests.
+The file is read from the PR head and is treated as untrusted data. It cannot choose commands, credentials, images, host paths, container privileges, or network policy. Tests are disabled by default; use `tests` mode only when the deployment explicitly accepts execution of repository code in the runner.
 
 ## Webhooks and deployment