Allow dev-command arguments and desktop port selection

[wesbillman]

Created by Brain on behalf of Wes.

Summary

Allow multiple Buzz development copies to select different frontend ports:

just desktop --port 1430
# In another terminal/worktree:
just desktop --port 1431 --no-watch
just web --port 1432 --host 127.0.0.1

• Forward arbitrary arguments from just web to Vite and just desktop to Tauri, using positional arguments rather than shell interpolation.
• A small desktop adapter translates --port N / --port=N into matching Tauri devUrl and Vite beforeDevCommand settings. Tauri's own --port controls its static-file server and would not configure Vite.
• Preserve the no-argument defaults, strict desktop port binding, and all arguments after --. Generated port configuration is prepended, before Tauri’s implicit runner-argument boundary. Explicit user --config values merge afterward and can override it. Runner-owned --port arguments need an explicit -- boundary.
• Document parallel copies and the distinction between ports and existing credential/plugin-profile sharing. No app-runtime, dependency, or profile lifecycle changes.

Validation

Checked at 982e7cd52f5ffff5350c6771a993292cf14123b9:

• bin/node --test tests/integration/dev-commands.test.mjs: 15 tests passed, executing the real just recipes and desktop adapter against a recording package-manager boundary. Includes default behavior, both port syntaxes, boundaries, invalid input, quoting, config order, and runner/application argument pass-through.
• With the base justfile substituted in an isolated temporary fixture, the same coverage fails for the newly requested behavior; the no-argument defaults still pass.
• Local smoke: two real just web --port ... servers and two real just desktop --port ... --no-watch --runner <stub> frontend servers responded concurrently. Captured Tauri's effective config confirmed matching port/URL, correct parsing even with implicit runner arguments, and preservation of an unrelated productName override.
• Occupied desktop frontend port failed without disrupting the original server. just desktop --help forwarded successfully.
• Pre-commit formatting/lint passed. Pre-push hooks selected no JS unit or design-system inputs; broad suites are left to existing CI. The new integration file is selected by CI's node --test tests/integration/*.test.mjs.
• Independent review: Pinky found an implicit-runner argument ordering defect in the first adapter; fixed with a prepended config and regression coverage. Correction review reports no remaining blockers.
• No browser test cases added or removed: this is a command/argument contract tested at the process boundary.

Native compilation and actual simultaneous native windows were not exercised: the smoke used the real Tauri CLI and Vite but substituted only the native runner. No live identity/relay activity. Full CI and code-owner approval remain separate gates.

Originating Buzz conversation: channel 3428ec3f-a58b-429b-afbc-dc6a77918ce8, thread 0470bf23405c699a93bba2d83a3791cfc50e6a1d5d63205761e1cb191a640094.

CI follow-up at 8e4e74acc39a234c475d60f046aca09d1e85c974

• Merged main at 9053fa9, incorporating test(browser): scroll past tall rows before asserting a new reading row #110. All four browser-journey jobs then passed, including the previously failing reading-anchor assertion.
• That run instead failed terminal::tests::bounded_input_and_session_admission during first PTY spawn with “nul byte found in provided data.” Its Node integration step was skipped. Failed Rust job.
• With Wes's authorization, repaired only src-tauri/src/terminal/tests.rs: both shell fixtures clone one OnceLock-initialized command with explicit HOME/SHELL. Locked portable-pty 0.9.0 otherwise calls non-reentrant getpwuid during construction when parent SHELL is absent and at spawn when HOME/SHELL are omitted. The template prevents concurrent fixture constructor lookups; explicit values avoid spawn fallback lookups. No production changes, serial test mode, retries, or weakened assertions.
• Added a fresh-process regression without parent HOME/SHELL: eight barrier-started workers each own a separate terminal registry, assert explicit command environment, and verify exact HOME/SHELL/PWD through real PTYs.
• Local macOS arm64: bin/cargo test -p buzz-foundation --lib --locked terminal::tests -- --nocapture passed all 11 terminal tests (3 unrelated native tests filtered out). Cold build 38.68s; test execution 0.73s. This is focused validation, not a full package-suite claim; CI owns broad coverage.
• Negative control in an isolated clone with separate Git metadata and build output: restoring both old fixture constructors while retaining the regression fails deterministically on missing explicit HOME. Restoring the candidate produces a clean tree and 11/11 passing terminal tests (0.75s). This establishes regression sensitivity, not reproduction of the historical intermittent NUL corruption.
• Pinky's independent read-only review at 8e4e74a / tests blob 4e7456e8ac81477dd81ba758b331a729ac97bd72 reports no blocking findings. Formatting and diff checks passed; mandatory hooks passed, with no JS/design tests selected by pre-push.
• New Linux CI results remain a separate gate. Actual native windows remain untested.

Refreshed integration head 8d5afc5726f4df8d4f22c3db9615711afb379f82

Main advanced during the repair and introduced a contributing.md conflict, which prevented PR CI from starting at 8e4e74a. Merged main 0b73a45, preserving both port documentation and local-notification guidance. The terminal fixture blob is unchanged. At this refreshed head, all 11 terminal tests (8 unrelated native tests filtered) and all 15 dev-command integration tests pass locally. That CI run completed: Rust/tool integration passed, including the previously failing admission test; Chromium shard 1 instead failed the agent-activity tooltip teardown/layout assertion described below.

Tooltip teardown repair at a5eb7da9eeb72c069f48937dfe53acda6dbcf5ab

• Saved Chromium trace identifies the exact 910px result: after Escape the closed tooltip portal remained at x=590 with a 320px popup when the viewport became 390px. Base UI 1.8.0 awaits animation-frame completion before unmount; key delivery is not completion.
• Five test-only lines await actual tooltip DOM removal (including hidden elements) before the next resize/click. Both immediate exact-width assertions and all geometry checks remain unchanged. No sleeps, retries, timeout increases, production CSS changes, or browser cases added/removed.
• Full affected agent-activity.spec.mjs passed in Chromium and WebKit: 10/10, 21.4s locally. Exact test SHA256: 40bcd0cfcd49f7882e727bac9a887b8c78b493265a614554ae0981b67202f7e2. This is focused browser validation, not the full repository suite.
• Controlled local diagnostic held the actual tooltip close-completion promise: Chromium without the removal barrier reproduces 910px; releasing teardown restores 390px. Removal-before-resize passes in both engines. WebKit did not reproduce overflow without the barrier (reported 390px), so cross-engine failure reproduction is not claimed. Diagnostic was removed, not added to the suite.
• Pinky's independent read-only review found no blocking issue. Mandatory hooks passed; DCO is green. Fresh hosted CI is running, not yet a passing result. Simultaneous native-window acceptance remains unverified.

[morgmart]

Integration update: #111 is notification-only. Dock labels are now implemented in draft PR #113, stacked on this PR’s shared desktop launcher. It preserves this argument/port interface and inherited package-manager/termination path; content-keyed icons fix warm-build invalidation. No code changes are requested here for Dock labels. After #109 merges, #113 will need retargeting to main and revalidation. The mistakenly created tracker #112 is closed; #113 is the implementation.