Testing and evidence

Run the native gates before each commit with Rust 1.90.0:

bash scripts/native-ci.sh verify
bash scripts/native-ci.sh release
bash scripts/native-ci.sh crate

verify checks release documentation labels, formatting, all-target compilation, Clippy and all-target tests. make verify remains the Docker alternative; Linux container results describe Linux, even when Docker runs on a Mac.

The October audit uses local Jenkins for CI, with immutable source selection and actual agent-platform validation. Automatic Actions jobs skip this audit branch; the optional four-target candidate workflow is manual-only. Do not trigger it without the owner's explicit request. Keep native-agent, Docker VM and unavailable platform results distinct in the acceptance record.

Real PTY regressions

tests/release_native_smoke.rs starts the executable under a real Unix PTY at 80×24 and other fixture sizes. It exercises editing, selection, Unicode paste, save, dirty-file protection, crash recovery, terminal restoration, compact layout and mouse/keyboard interactions. Its AI security fixtures use loopback mock providers and dummy credentials: a hostile shell response cannot create a marker, and a Claude redirect cannot contact the second endpoint with its key. These are transport and editor regressions, not live-provider qualification.

cargo test --locked --test release_native_smoke
MELLOW_SMOKE_BINARY=/absolute/path/to/extracted/mellow \
  cargo test --locked --test release_native_smoke -- --nocapture

The second command tests exactly the selected executable. Source unit tests do not establish archive acceptance. native-ci.sh artifact-install extracts a version/target-matching release archive, verifies the executable version, runs all PTY regressions and writes archive/binary SHA-256, source commit, host and outcome to an adjacent .acceptance.txt file. A failed or unexecuted test fails the gate. Native macOS Intel, macOS ARM64, Linux x64 and Linux ARM64 are separate targets; emulated runs do not satisfy a native artifact gate.

Manual terminal and integration evaluation

Open fixtures/unicode.txt and verify Telugu, CJK, emoji, combining marks and tabs with actual fonts. Exercise Terminal.app, iTerm2, Ghostty and Linux/WSL terminals; SSH with and without tmux; legacy and enhanced keys; 16/256/true color; bracketed paste on/off; mouse and keyboard-only navigation. Record terminal version, OS/architecture, font, transport and exact executable digest. PTY text parsing does not establish native glyph shaping or screen-reader usability.

Live Python/TypeScript language-server completion/imports, mixed-language navigation/diagnostics and live AI provider/account compatibility require their own named integration results. Mocks establish only their protocol fixtures. VoiceOver or other screen-reader acceptance requires a real assistive-technology session and user evaluation; keyboard/color tests alone are insufficient.

Evidence status

The automated gates are implemented; their presence is not a passing run. Release evidence must link the actual logs and archive digests for each target. Native Intel/Linux jobs, real SSH/tmux, live provider accounts, native font rendering and screen-reader/user evaluations remain unverified until recorded for the candidate. See the implementation tracker for current local results.

Manual Jenkins job

Create a manual Pipeline using the reviewed Jenkinsfile, without SCM polling. SOURCE_SHA is mandatory and immutable; use the public Mellow repository or an absolute agent-local checkout for SOURCE_REPOSITORY. The default agent is the existing vidhur-mac && darwin-arm64, with command-scoped Rust 1.90.0. The job does not change global toolchains, credentials, other jobs or installed Mellow. When the controller root URL is unset, JENKINS_BASE_URL supplies the evidence URL without changing controller settings. It runs formatting/lint/source tests, release/crate, isolated installation, and all PTY tests against the extracted native archive. Optional Linux VM packages are labeled separately and include deb/rpm install/test/uninstall. Failure logs and source/archive/binary receipts are always archived.

Only enable optional Intel or native Linux gates when a correctly labeled agent exists. Host/toolchain/target mismatches fail before product tests. A timeout or missing agent is an unavailable gate, never a passing target. Check the actual Jenkins build result and artifact receipts linked in PR #11.