test(formats): port the Open VSX gateway probe as test-vscode-remote (release gate)

[cazlo]

fixes #375

Summary

artifact-keeper/artifact-keeper#3253 merged 2026-08-15, shipping AK as a governed VS Code extension gateway (composed Open VSX gallery pull-through, every asset URL rewritten onto AK). The PR's own dogfood harness carried a note that its probe subcommand should move to the release gate; it never did. This adds tests/formats/test-vscode-remote.sh, ported from scripts/dogfood-vscode-openvsx.sh's probe subcommand, and registers it in both format-tests.yml and release-gate.yml (node batch in each).

What it covers

• Gallery manifest names AK's own extensionquery/latest endpoints, not open-vsx.org's.
• extensionquery POST with the exact VS Code client request-body shape resolves a reviewed extension (redhat.vscode-yaml) and every returned assetUri/fallbackAssetUri/files[].source is AK-owned.
• {publisher}/{name}/latest resolves the same extension, all asset URLs AK-owned.
• The code-server compatibility alias (/gallery/vscode/{publisher}/{name}/latest) resolves identically to the primary latest lookup, checked as its own direct probe so a regression in the alias specifically can't hide behind the primary path passing.

Not ported: the dogfood script's vscodium-ui, code-server-ui, install-, update-, and TRACE_NETWORK subcommands drive a real VS Code-family client against an isolated profile. That doesn't fit a headless, 120s-per-script release-gate suite, and the gate runners have no VSCodium/code-server binary on PATH.

I also fixed a real bug while porting rather than copying it forward: the dogfood script's inline jq asset-URL assertions build [.assetUri?, .fallbackAssetUri?, (.files[]?.source?)] without filtering nulls before all(startswith($prefix)). The ? suffix suppresses type errors, not missing keys, so a version object lacking fallbackAssetUri still yields null for that slot and null | startswith(_) is a jq runtime error, which aborts the whole jq -e and inverts the diagnosis (reads as "retains a non-AK asset URL" when the truth is "field absent"). The new _assert_asset_urls_ak_owned helper null-filters first.

Validation against the shipped composed gateway

artifact-keeper/artifact-keeper#3253 landed a gallery skeleton+synthesis rework on 2026-08-15, after I first wrote this port, so I re-validated every assertion against a live composed-gateway backend rather than trusting the pre-rework version of the script.

Bounded extensionquery responses (I measured 30 stable + 29 prerelease versions returned for the probe extension, against the newest-30-per-platform-x-channel cap, instead of raw upstream counts), the /latest age-gate walk-back (200 with the newest allowed version instead of 404), and the flags-conditional presence of files/assetUri on synthesized entries are all confirmed compatible with this script as written. None of the assertions check an exact or all-versions count, so the bounding doesn't reach them, and the script always queries with flags=511, which requests full asset detail, so it's never exposed to the omitted-field case. No assertion needed changing.

I then ran the script end to end against that backend: 9 passed, 0 failed, 0 skipped in 3s, against the gate's 120s-per-script budget. The run also confirms the cleanup path — the throwaway repo it creates is gone afterwards — and that the backend accepts format: vscode with an Open VSX gallery-adapter upstream_url through the ordinary repository-creation API.

Red-proof

This is new-feature coverage, not a regression test for an existing defect, so there's no prior "broken behavior" to construct a red state against. The natural red state is that these routes didn't exist at all before artifact-keeper/artifact-keeper#3253 merged (404, not a wrong-answer 200) -- I'm treating route-absence as sufficient rather than adding a synthetic failure-injection leg, consistent with how this repo's other remote-proxy format tests (npm-remote, go-remote, etc.) were added as straight coverage additions without a constructed red leg.

Also found while here

Having to register this script by hand sent me looking at what else isn't registered. tests/formats is the only suite without auto-discovery — fourteen other suites run through scripts/run-suite.sh, which globs tests/<suite>/**/test-*.sh, while formats enumerates scripts by name in the workflow YAML. That gap leaves 59 of the 119 tests/formats/test-*.sh scripts named in no workflow at all, including a regression gate written for artifact-keeper/artifact-keeper#2477 that has never executed. Filed separately as #374 rather than folding a repo-wide fix into this PR.

Test plan

• Manual run against a live composed-gateway backend: 9 passed, 0 failed, 0 skipped, 3s.
• format-tests.yml node batch, release-gate.yml node batch: both green in this repo's own CI.