ci: add http + unix VGI transport coverage to SQL E2E

[rustyconover]

Adds subprocess + http + unix VGI transport coverage to the SQL E2E suite, with the full test/sql/*.test suite GREEN over all three transports.

What changed

• ci/run-integration.sh is now parameterized by TRANSPORT (subprocess | http | unix, default subprocess). The mock MediaWiki server stays up for all transports, and VGI_WIKIPEDIA_API_URL is exported into the environment (not baked in as an env VAR=… <cmd> prefix) so both the DuckDB-spawned subprocess worker and the out-of-band http/unix workers inherit it.
  • http: injects INSTALL httpfs FROM core; LOAD httpfs; into staged tests (the vgi HTTP transport rides DuckDB's httpfs), boots wikipedia_worker.py --http --port 0 --port-file <f>, polls the port-file, ATTACHes http://127.0.0.1:<port>.
  • unix: boots wikipedia_worker.py --unix <sock>, polls for the socket, ATTACHes unix://<sock>.
  • Silent-skip guard: the sqllogictest runner auto-SKIPs HTTP errors (fake green); the leg fails unless the runner reports All tests passed (N>0 assertions) with zero skips.
  • One cleanup trap captures $? first and kills both the mock server and the worker.
• CI (.github/workflows/ci.yml): the integration job is now a transport: [subprocess, http, unix] × os: [ubuntu-latest, macos-latest] matrix. Test + lint jobs unchanged.
• Python http extra: added http = ["vgi-python[http]>=0.8.3"], listed vgi-python[http] in the worker's PEP 723 header, relocked, and uv sync --frozen --extra http in the integration job.

Streaming over HTTP

wiki_search's sroffset paging already externalizes its cursor into a serializable ScanState, so it round-trips correctly over the stateless HTTP transport — the full suite (incl. wiki_search.test's max_pages := 2 batch/page-boundary test) runs over http with no gate and no worker change.

Local validation

• subprocess: GREEN — 51 assertions
• unix: GREEN — 51 assertions
• http: GREEN — 55 assertions (51 + injected httpfs INSTALL/LOAD)

🤖 Generated with Claude Code