CI: gem variant apply tests assert JSON codes (#1089) - #1642

LLM Description written by Claude Code:claude-opus-5-5

Refs #1089 (the apply gem-variant slice of child 6; the tracking issue stays open)

Root cause

apply/cli_gem_variant_mismatch_policy.rs checked gem release-variant
behaviour by matching 8 English sentences on human stderr. A wording
change could fail a test whose behaviour was still right. The rest of
the apply suite already asserts --json codes in its behaviour
tests and keeps the wording in apply_render.rs.

Change

• cli_gem_variant_mismatch_policy.rs: all 6 tests now run with
  --json and assert the exit code, status, event actions and
  errorCodes, plus the bytes on disk:
  • default overwrite: applied plus a content_mismatch_overwritten
    event for the purl.
  • --strict: failed/apply_failed for the purl, naming the
    file, and no overwrite event.
  • --force: applied plus the overwrite event. Before this, only
    the exit code and bytes were checked.
  • wrong-platform qualified singleton, and a multi-variant base where
    no variant matches: partialFailure with no applied, failed
    or overwrite event, and the file untouched.
  • mismatched sibling: only the installed variant is applied, with
    no overwrite event.
• apply_render.rs: 5 new render tests hold the human wording:
  the mismatch warning line (checked for the purl on that line), the
  --strict hash refusal, "no matching variant found" for each of the
  two no-match cases, and no overwrite warning for a skipped sibling.
• apply_fixture.rs: now holds the gem variant fixtures
  (run_gem_apply with BUNDLE_* scrubbed, and the singleton,
  wrong-platform and multi-variant fixtures). They reuse its existing
  install_gem, patch_record, write_manifest and stage_blob,
  which drops the policy file's private copies.

Assertions of 4+-word sentences in cli_gem_variant_mismatch_policy.rs
went from 8 to 0. No production code changes.

Testing

• cargo test -p socket-patch-cli --all-features --test apply: 153
  passed, 0 failed, 1 ignored. Main has 148: the 6 policy tests stay,
  and 5 render tests were added.
• cargo fmt --all -- --check: clean.
• cargo clippy -p socket-patch-cli --all-features --test apply -- -D warnings
  reports nothing in the three changed files. Its other findings are
  already on main, in tests/common. CI's clippy command doesn't build
  tests, so it's unaffected.
• The expected JSON shapes come from running each fixture with
  --json on main's binary.
• CI on a641164 is green: 36 passed, 50 skipped (full-tier jobs), no failures. Bugbot found no issues.
• No cfg(windows)/cfg(unix) code changed, so I didn't dispatch a
  full-scope run.

CI cost

Test-only change in an existing test binary. A lean PR run is the same
before and after: about 28 Linux jobs, about 8 minutes.

/code-review high findings

• Fixed: the fixtures moved into apply_fixture.rs instead of being
  imported from a sibling test file. The no-match tests now check for no
  failed event. The --strict event must name the file. --force
  now checks the envelope. The render purl check is on the warning line
  itself. The two no-match render cases are now separate tests. No
  constants are exported without a user.
• Not fixed here (product JSON contract, outside this test lane): in
  --json, a base with no matching release variant exits 1 with
  status: partialFailure and summary.failed: 0. Its variants show
  only as skipped/package_not_installed, and no event names the
  base or a no_matching_variant code (commands/apply.rs around
  lines 2466-2479 and 3040). A JSON consumer can't tell which package
  is still unpatched. The tests pin today's shape, so a fix that adds a
  failed event will update them deliberately.

Generated by Claude Code