Add a build-dstack job to docker.yml that builds with
CARGO_FEATURES=dstack and tags as :dstack / :dstack-<sha> /
:<version>-dstack. Only runs on push to main and releases.

Fix docker-compose.dstack-postgres.yaml:
- Use :dstack image tag instead of :latest
- Add restart: unless-stopped
- Remove port 8001 (admin via phala ssh only)

The dstack API changed its /GetKey endpoint response from
`{"asBytes": "<hex>"}` to `{"key": "<hex>"}`, causing a
deserialization crash on Phala Cloud.

The scratch-based Docker image has no CA certificate bundle, causing
`no CA certificates found` panics when making HTTPS connections to
external services (Neon Postgres, Hetzner S3). Copy the cert bundle
from the Alpine builder stage.

…37)

S3-compatible providers (Hetzner, etc.) require a region param.
Add AWS_DEFAULT_REGION from env file. Revert pinned SHA tag back to :dstack.

The scratch base image has no /tmp directory, causing "No such file or
directory" errors on write operations that use temp files. Create an
empty /tmp with sticky bit in the builder and copy to runtime.

* fix: add /tmp directory to scratch runtime image

The scratch base image has no /tmp directory, causing "No such file or
directory" errors on write operations that use temp files. Create an
empty /tmp with sticky bit in the builder and copy to runtime.

* fix: set TMPDIR=/data in Dockerfile for scratch runtime

The scratch base image has no /tmp directory and no writable temp location.
Set TMPDIR to /data which is owned by the tinycloud user and always exists.

* fix(core): flush in-memory SQL databases to file on actor shutdown

SQL database actors start with an in-memory SQLite instance and only
promote to file storage when data exceeds the memory_threshold (10 MiB).
Small databases never hit this threshold, so when the actor shuts down
after the 5-minute idle timeout, all data is silently lost.

This adds a flush step just before the actor exits: if the database is
still in-memory and contains any data, it is promoted to file via the
existing SQLite backup API. On the next request, the actor sees the file
and opens from disk — data survives across idle cycles.

KV is unaffected because it writes to the server's main persistent
database immediately. This only impacts the per-space SQLite service.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* chore: add changeset for SQL flush fix

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>