Qt-removal R7.5c: global keyboard shortcuts

[dovvnloading]

Ports legacy's QShortcut block (graphlink_window.py:307-318) and its AcceleratorForwardingFilter typing arbitration (graphlink_web_island_host.py:692-755) into the SPA. Third of the R7.5 parity sub-increments, after R7.5a (mechanical gaps, #147) and R7.5b (canvas visuals, #148/#149/#150).

What lands

Eleven bindings, all matched against legacy 1:1:

Binding	Action	Gated while typing?
Ctrl+T	New Chat (with confirm)	yes
Ctrl+L	Toggle Library	yes
Ctrl+S	Save Chat	no
Ctrl+K	Command Palette	no
Ctrl+F	Search	yes
Ctrl+G / Ctrl+Shift+G	Create Frame / Container	yes
Ctrl+Arrow ×4	Branch navigation	yes

The two exemptions are legacy's own, and legacy contract-tests them in test_keyboard_arbitration.py. The suppression set here is exactly those nine combinations, mirrored in a test so a future edit has to be deliberate rather than incidental.

Two new pure modules keep the logic unit-testable without mounting <ReactFlow>, the same posture smartGuides.ts established in R7.5b-3:

• chrome/shortcuts.ts — resolveShortcut key matching plus isGatedWhileTyping. Requires Ctrl-or-Cmd, rejects Alt so AltGr-produced characters still type, and rejects Shift on bindings legacy never defined with it (Ctrl+Shift+G is a separate binding, not a modifier variant of Ctrl+G).
• canvas/treeNavigation.ts — resolveTreeNavigationTarget, porting _navigate_up/_down/_left/_right. Up goes to the parent; down to the leftmost child by x-position, not the first-created; left/right step through x-ordered siblings and require a parent; every boundary is a pure no-op with no wrap. Only chat/conversation/html kinds are navigable, which reproduces legacy's children graph — it never contained code/document/image/thinking nodes, even though this backend makes all of them real edge targets.

App.tsx's GlobalShortcuts grows from two ungated bindings to the full dispatch. requestNewChat restores legacy's blocking confirm, and a save-chat palette entry closes a genuine gap: Save has had an app-bar button and a store method since R6.5 but no palette command.

Three defects found and fixed before shipping

1. setCenter lost the node half-size. It used measured?.width ?? 0, which silently degrades to centering on the node's top-left corner — the exact empty-measured trap R7.5b-3 already diagnosed and solved, because toFlowNodes rebuilds the node objects RF derives measured from. measuredNodeSize is now exported from SceneCanvas and reused rather than reimplemented badly.

2. requestNewChat erred in the unsafe direction. Legacy skips its confirm only when the canvas is empty and there is no current chat id; this skipped on emptiness alone. So emptying a loaded chat and pressing Ctrl+T detached it with no prompt — newChat() drops current_chat_id, so the next Save would INSERT a new row instead of updating the chat the user thought they were editing. The in-code comment had claimed the divergence erred safe; it erred the other way. Fixed by publishing hasSavedChat, the one derived bit of current_chat_id the frontend needs to evaluate the real predicate. The row id itself stays server-side, as that field's own comment has always specified.

3. Branch navigation worked for exactly one hop. This is the one the automated suite structurally could not catch, and it broke the feature outright. setCenter fires onMove → onMove reports the viewport → the backend echoes a fresh scene → toFlowNodes rebuilds every node object, wiping the selection the keystroke had just made about 300ms earlier. The next Ctrl+Arrow then found no single selected node and no-opped, so you could never walk more than one step.

That wipe is pre-existing, not introduced here: any selection that merely overlaps a snapshot — an autosave tick, a streaming token — loses its highlight the same way. R7.5c is simply the first feature that depends on selection surviving. It is therefore fixed at the source, with a new exported pure withPreservedSelection(rebuilt, current) re-applying selection across the rebuild, covered by five unit tests including one asserting a backend-deleted node cannot be resurrected by a stale id.

Test plan

• 868 backend tests (pytest backend/ tests/test_no_qt_anywhere.py)
• 1201 frontend tests (npm run check: schema → typecheck → lint → test → build), 0 lint errors
• Burn-down gate unchanged at 152/84/68 — pure feature work, no Qt files touched
• Codegen re-run; generated TS/schema in sync with the payload dataclass
• Adversarial review of the diff (found defects 1 and 2)
• Live verification against a real backend with real KeyboardEvents and a WS-seeded 4-node branch scene:
  • Ctrl+Down selected the leftmost child by x (200), not the first-created (600)
  • Chained Ctrl+Left walked x=900 → 600 → 200 across separate keystrokes — the case that was broken
  • Left boundary was a clean no-op; Ctrl+Up returned to the parent
  • Every hop centered on the node's exact geometric center (e.g. 1111,456 = 900+422/2, 400+112/2), proving fix 1
  • With the composer textarea focused, Ctrl+Down/Left/L/G/T were all suppressed — no selection change, no view change, no overlay, no node created
  • Ctrl+K still opened the palette with that same textarea focused, confirming the exemption
  • Ctrl+G with exactly one node selected created a real frame, confirming the shortcut's 1+ gate

Documented divergences

Both deliberate, both pre-existing, neither silently changed here:

• The palette's Create Frame/Container keep their stricter 2+ selection gate while the shortcuts use legacy's 1+. That gate is an R6.1 choice; changing shipped palette behavior inside a shortcuts increment would be scope creep, so it is flagged rather than altered.
• Ctrl+K and Ctrl+F remain toggles where legacy was show-only (an R2.4 divergence). Only their focus-gating was corrected here — Ctrl+F was previously ungated and should not have been.