Bump sqd-query-engine to 6df565a, version 2.16.0-rc.1

[define-null]

Summary

Bump sqd-query-engine (the experimental dynamic query engine) from 54b9020 to 6df565a, the current upstream master tip, and set the worker version to 2.16.0-rc.1. The previous pin dated from the initial integration in July and had been orphaned upstream (not reachable from master).

Upstream range: 52 commits. Full list: subsquid/sqd-query-engine@54b9020...6df565a

The catalog format changed

Two of those commits reshape the dataset description the engine loads, and DatasetDescription is deny_unknown_fields, so a stale catalog is a load error rather than a silently ignored key:

• c31a4a0 makes filters a required, closed list — a table that accepts none says filters: [].
• b72c404 groups a table's keys into request (what a client may send), output (what a client may see) and the storage keys. query_name → request.name, field_name → output.name, json_encoding: hex → encoding: hex_bytes; stats, dictionary, children and parent_key are gone.

Both surfaces are now closed: a column is filterable only if request.filters names it, and selectable only if output.fields does.

The four inlined test catalogs still spoke the flat layout, which is what the first CI run caught — 23 failures, all of them missing field 'filters' or an unknown key. They are rewritten here; no production code needed changes.

Note for whoever ships this: the catalogs the worker fetches at runtime have to be republished in the new layout too, or the worker hits the same load error outside the tests.

One bad schema no longer takes the bundle down

Separate commit. Parsing a staged schema failed the whole bundle, which is how a stale catalog for
a kind this worker holds no chunk of would stop it serving the kinds it does hold. load_dir now
drops the schema it cannot parse, warns, and counts it (schema_bundle_schemas_skipped), leaving
the rest installed.

That is not a loosening: the id is left out of bundle_ids, and DatasetsIndex already refuses an
assignment whose chunks reach for an id the bundle does not offer — so the refusal lands on the
assignment that actually depends on the schema rather than on every assignment. A bundle in which
nothing parses stays a permanent fault, since no query could be served from it. The store path had
this shape already: read_store skips a file it cannot read and lets a later bundle reinstall the
id.

Test plan

• cargo check --all-targets --locked
• cargo test --all-features --lib --bins --test state_pbt --test state_regression
• Conformance tier (e2e, query_surface, query_concurrency), default and --all-features
• New: a bundle carrying one unparseable schema installs the other two and offers only their ids
• New: a bundle in which nothing parses is still refused as permanent, with nothing left in the store

🤖 Generated with Claude Code

https://claude.ai/code/session_01MEHFzAGrwxmu429ck2oDk9