fix: Iceberg lookup double-encodes user_meta (0.19.1)

[drernie]

Fixes #399.

The bucketless Iceberg lookup projected the metadata column as json_format(CAST(m.metadata AS JSON)), but m.metadata is a VARCHAR that already contains a JSON document — and Trino's VARCHAR→JSON cast wraps the string as a JSON string value instead of parsing it. Every row returned by the Iceberg path therefore came back double-encoded: json.loads yielded a str, _parse_user_meta logged "Athena metadata JSON was not an object", and the result's metadata silently degraded to the matched {key: value} fallback. Deterministic on every lookup, first observed during live verification of quiltdata/deployment#2502 on tf-dev-bucketless.

Canvas rendering was unaffected (it only consumes bucket + package name), so this is log noise plus a latent trap for future metadata consumers — hence a patch release, not a re-pin of the 0.19.0 deployment.

Changes

• _build_iceberg_union_query: project m.metadata AS user_meta raw, matching the legacy _packages-view path (which selects user_meta directly and never had this problem). The WHERE json_extract_scalar(...) filter was always correct — json_extract_scalar does parse VARCHAR — which is why matching worked while the projection double-encoded.
• _parse_user_meta: if the first decode yields a str, defensively decode once more before warning, so any double-encoded values still round-trip to the full dict.
• Tests: the union-query test now asserts the raw projection (it previously asserted the buggy json_format(CAST(...)) form); new test covers double-encoded user_meta round-tripping to the full metadata dict.
• Version bump 0.19.0 → 0.19.1 + changelog.

Test plan

• docker/: 468 Python tests pass, including the new double-encode round-trip test and test_versions_match
• Existing non-object-metadata fallback test ('["EXP-1"]') still passes — genuinely non-object metadata keeps the warning + fallback behavior

🤖 Generated with Claude Code

Greptile Summary

This PR fixes Iceberg lookup metadata decoding for bucketless package searches. The main changes are:

• Projects Iceberg m.metadata as raw user_meta instead of wrapping it with json_format.
• Adds a defensive second JSON decode when user_meta arrives double-encoded.
• Updates Iceberg lookup tests for raw projection and double-encoded metadata recovery.
• Bumps the app, Python package, lockfile, and npm package versions to 0.19.1.
• Adds the 0.19.1 changelog entry.

Confidence Score: 5/5

This looks safe to merge.

• No blocking issues found in the changed code.

Important Files Changed

Filename	Overview
docker/src/package_query.py	Changes Iceberg lookup projection to return raw metadata and adds double-encoded JSON recovery.
docker/tests/test_package_query.py	Updates Iceberg SQL assertions and adds coverage for double-encoded metadata parsing.
CHANGELOG.md	Documents the 0.19.1 metadata decoding fix.
docker/app-manifest.yaml	Bumps the app manifest version to 0.19.1.
docker/pyproject.toml	Bumps the Python package version to 0.19.1.
docker/uv.lock	Aligns the locked editable package version with 0.19.1.
package.json	Bumps the npm package version to 0.19.1.

_{Reviews (1): Last reviewed commit: "docs: changelog for 0.19.1" | Re-trigger Greptile}