fix(gpui): unbind macOS ⌘ shortcuts and wire ⌘Q to quit

[zhbjerry]

Result / 用户结果

Fixes #238.

On macOS the shell's ⌘ shortcut set lived outside the keybinding model, so it could not be released, and Action::Quit was never wired at all. Now:

• Clearing an action releases its ⌘ key together with its configured key. Clearing the Shell picker row writes keybind=Ctrl+K:ReceiveChar and keybind=cmd+k:ReceiveChar, and ⌘K stops opening the panel immediately, without a restart. Restoring the row brings both keys back.
• ⌘Q quits through the same path as tray quit: drafts and session are saved before PTYs stop.
• On macOS both the keymap page and the lines the app writes into pebrel_settings.txt spell the super modifier Cmd+ instead of Win+, so Ctrl+Win+F / Win+F are gone.

Users upgrading from a released build must clear the row once more: their existing file predates the alias vocabulary and holds no line for the ⌘ half. Nothing else changes for them — win+ / super+ / cmd+ stay equivalent in keybind= lines, and the settings page continues to show the configuration table's keys.

Out of scope / 不在本次范围

Unbinding ⌘K does not make it clear the screen, and this PR does not add that. The key is released, not re-bound to anything:

• The GPUI terminal encoder translates Ctrl, Alt and named-key combinations and returns None for a pure ⌘ combination (nebula_app/src/gpui_shell/terminal/keymap.rs:276), so nothing reaches the PTY.
• Pebrel's own macOS default table already declares ⌘K as clear screen (nebula_app/src/config/bindings/defaults.rs:367: Esc("\x0c") plus ClearHistory), but workspace_binding_in_context has no arm for Action::Esc / Action::ClearHistory, so the shell never implemented it.
• Doing it properly means adding a clear-screen action (write 0x0c, clear scrollback) and deciding whether macOS ⌘K should belong to that action instead of the Shell picker — a default-behavior change. I can submit it separately once the maintainer agrees on scope.

Ctrl+L still passes through to the shell (defaults.rs:81 keeps it as ReceiveChar); it clears the screen at the shell level but not the app's scrollback.

Design / 设计边界

• Responsibility and affected modules: display::keymap stays the single authority for defaults, unbinding and restoring; gpui_shell/workspace/keyboard_bindings.rs only translates that table into gpui bindings. MACOS_COMMAND_ALIASES replaces the second hand-written ⌘ list in bind_macos_command_keys — keeping two copies is what produced the defect.
• Why this belongs here; interfaces that remain unchanged: default_shortcuts() gains the alias entries on macOS, deduped against config-table entries that already cover the same action and key. Entries are written verbatim into keybind= lines because the reload path matches a restored default by string, so the table spelling is the persisted spelling (cmd+…, modifier order as in canonical_combo). The keybind= format, the default_shortcuts() signature, the settings page contents and the legacy shell are unchanged.
• Action::Quit maps to a new QuitApp action whose handler defers windowing::quit_all. It is registered in keyboard_bindings::init because nebula_app/src/gpui_shell/workspace.rs sits at its hard budget ratchet and cannot grow.
• Dependency, data-format, threading, or lifetime changes: none. The only new persisted lines are the cleared keys themselves.
• Compatibility and migration/fallback behavior: covered under Result; a user who wants the ⌘ half released again just clears the row once in this build.
• Decision record: architecture/notes/nebula_app/gpui_shell/2026-09-21-macos-command-key-aliases.md — single authority, rejected alternatives, and the ⌘K clear-screen scope note.

Evidence / 验证依据

• Commands and actual results (all on macOS 26 arm64, pinned toolchain, --locked):
  • cargo fmt --all -- --check → clean.
  • python3 scripts/check_architecture.py --base 3050374 → no errors; pre-existing NOTICEs only; no budget or allowance raised.
  • python3 scripts/check_platform_cfg.py → platform cfg outside platform/: 484 (budget 484) (unchanged).
  • cargo test -p nebula --bin pebrel --features gpui-shell,gpui_platform/runtime_shaders --locked display::keymap → 14 passed.
  • cargo test -p nebula --bin pebrel --features gpui-shell,gpui-test-support,gpui_platform/runtime_shaders --locked keyboard_bindings → 7 passed (includes the tab-dispatch tests).
  • cargo test -p nebula-settings --locked → 67 passed. cargo test --manifest-path tools/i18n-contract/Cargo.toml --locked → 20 passed, 1 pre-existing ignored.
  • python3 -m unittest scripts.tests.test_architecture_budgets test_architecture_dependencies test_architecture_governance → 53 passed; ... test_pebrel_branding test_windows_resources test_app_icons test_ai_session_probe → 18 passed, 1 skipped.
• Regression tests: what fails before the fix?
  • clearing_releases_the_macos_command_alias fails without the change with ⌘K 必须按表里的写法逐字写回，实际 [("Ctrl+K", "ReceiveChar")] — i.e. the settings write only releases Ctrl+K, exactly the reported bug.
  • macos_command_aliases_round_trip_through_storage pins that every alias entry parses back to the same combination, so a cleared key is never written as an unparsable line.
  • every_macos_command_alias_binds_a_gpui_action and released_cmd_k_is_swallowed_by_no_action pin the shell wiring: the table has a binding, and the injected no-action binding suppresses the static default at Root and terminal depth.
  • super_modifier_renders_per_platform pins the per-platform keycap.
• UI changes: screenshots, long translations, keyboard access, DPI checks: manual acceptance with a debug build of this branch, driving the real window (isolated PEBREL_CONFIG_DIR): default ⌘K opens the Shell picker → clearing the row writes Ctrl+K:ReceiveChar + cmd+k:ReceiveChar and both ⌘K and Ctrl+K stop opening it without a restart → restoring the row brings ⌘K back → ⌘Q exits and session.json records clean_exit: true; the keymap page shows Ctrl+Cmd+F for Fullscreen (was Ctrl+Win+F). No layout, translation or DPI surface is touched by this change.
• Hot-path changes: none. default_shortcuts() runs only for the keymap page and keybind reload; the macOS branch adds one linear pass over a 25-entry table.
• Unrun checks: Linux and Windows suites were not exercised locally (the alias table is test data there; the runtime path stays behind cfg(target_os = "macos")). There is no end-to-end test for ⌘Q — its handler runs the asynchronous draft-approval and session-save path, so it was verified in the real window instead.

Required Review / 必须确认

• I followed CONTRIBUTING.md, docs/architecture.md, and docs/project-constraints.md; the decision is recorded under architecture/notes/.
• I split responsibilities, not arbitrary line ranges; no duplicate behavior authority was added — the duplicated macOS ⌘ list was removed and replaced by one table.
• python3 scripts/check_architecture.py --base 3050374 passes; budgets were not inflated to fit the change.
• Tests cover success and failure; platform/feature coverage limitations are stated above.
• New messages use typed i18n IDs and matching placeholders; untranslated content has an explicit fallback. (Not applicable: no user-facing message was added, comments only.)
• Governance changes include a counterexample, corrected contract, tests, and a maintainer-reviewed decision. (Not applicable: no policy, gate or contract was changed.)

Checkboxes explain the review; they do not replace CI or maintainer approval.

[zhbjerry]

CI status: the two red tests are pre-existing on main

Tests (ubuntu-24.04) and Tests (macos-26) fail on tests this PR does not touch, and the same tests already fail on main:

• ai_hook::integrations::tests::generic_agent_binaries_require_a_matching_installation — nebula_app/src/platform/agent_integrations.rs:178, Os { code: 21, kind: IsADirectory } in the tests::executable helper.
• gpui_shell::workspace::details_panel::tests::file_path_edit_navigates_and_keeps_the_last_directory_on_error_or_escape — nebula_app/src/gpui_shell/workspace/details_panel_tests.rs:78.

Evidence that they are not caused by this branch:

• main at 30503743 fails with exactly these two tests: run 35597825341 — ubuntu-24.04, macos-26 and macos-26-intel are all red with the same names.
• nebula_app/src/platform/agent_integrations.rs does not exist at this PR's base commit f7ca0ec (git cat-file -e f7ca0ec:nebula_app/src/platform/agent_integrations.rs → fatal: path ... does not exist). The file arrives with later main commits, so the failure comes from the merge with current main, not from the diff here.
• Another run on the internal branch fix/cwd-session-restore-20260921 (run 35596902833, head 927a2c42, merged as PR Integrate complete Pebrel 1.9.0 desktop and Windows session recovery #231) also fails on ubuntu-24.04 with the identical agent_integrations test.

Everything this PR changes is green in CI:

display::keymap::tests::clearing_releases_the_macos_command_alias ... ok
display::keymap::tests::macos_command_aliases_round_trip_through_storage ... ok
display::keymap::tests::super_modifier_renders_per_platform ... ok
gpui_shell::workspace::keyboard_bindings::tests::macos_command_keys::every_macos_command_alias_binds_a_gpui_action ... ok
gpui_shell::workspace::keyboard_bindings::tests::macos_command_keys::released_cmd_k_is_swallowed_by_no_action ... ok
gpui_shell::workspace::keyboard_bindings::tests::cleared_shortcut_reaches_terminal_and_can_be_restored_without_restart ... ok

Passing jobs on this head: architecture-contracts, Tests (windows-2022), Tests (windows-11-arm), Release workspace (macos-26), Release workspace (macos-26-intel). Tests (macos-26-intel) was still queued when I wrote this.

本地与 CI 的失败集一致，两个红色测试在 main 上同样失败，且其中一个所在的文件在本 PR 的 base commit 上还不存在；本 PR 触及的 display::keymap 与 keyboard_bindings 测试全部通过。

[zhbjerry]

Rebased onto main @ 3050374 and force-pushed; new head is 4d08497. The diff is unchanged (4 files, +312/−28) and re-verified after the rebase:

• cargo fmt --all -- --check → clean.
• python3 scripts/check_architecture.py --base 3050374 → no errors, pre-existing NOTICEs only.
• python3 scripts/check_platform_cfg.py → 484 (budget 484), unchanged.
• cargo test ... display::keymap → 14 passed; cargo test ... keyboard_bindings → 7 passed.

The two failing tests noted above (agent_integrations, details_panel) are now part of the checked-out base, so they are expected to stay red here until they are fixed on main.

[zhbjerry]

Settled CI on the rebase head (rebase onto main @ 3050374), run 35604522645:

job	result
architecture-contracts	pass
Tests (windows-2022)	pass
Tests (windows-11-arm)	pass
Release workspace (macos-26)	pass
Release workspace (macos-26-intel)	pass
Tests (ubuntu-24.04)	fail — agent_integrations.rs:178
Tests (macos-26)	fail — agent_integrations.rs:178, details_panel_tests.rs:78
Tests (macos-26-intel)	fail — agent_integrations.rs:178, details_panel_tests.rs:78

The three red jobs fail only in those two tests, both untouched here and both already red on main @ 3050374 before this branch existed. No test under display::keymap or gpui_shell::workspace::keyboard_bindings fails on any platform.

[zhbjerry]

Closing this in favour of a replacement PR opened from a rebuilt branch. The diff and the commits are unchanged (4 files, +312/−28); only the branch history differs.