emrg: gui — restore stop button to interrupt active reply

[argszero]

Restores the send↔stop toggle that the React migration dropped (rant 2026-09-02T20:30:05): the GUI had no way to interrupt an in-progress reply while the TUI supports ESC interruption.

Changes:

• Composer (renderer): when busy, the send button is replaced by a red stop button (■, composer.stop tooltip/aria-label) wired to window.emrg.cancel (preload bridge already exposed; daemon cancel + releaseOwnStream untouched).
• Esc shortcut: with the command menu closed and the composer busy, Esc triggers the same stop (TUI muscle memory); with the / menu open, Esc still only closes the menu (no accidental stop).
• Injected cancel prop on ComposerProps (default window.emrg.cancel, aligned with the sendMessage/saveImage injection style) for tests.
• Stop optimistically clears the typing indicator and appends a system message ⏹ 已停止响应。 / ⏹ Interrupted — response stopped. (new chat.interrupted key, zh+en, 394 keys aligned). The daemon's cancelled broadcast then idempotently releases the external busy lock (existing releaseOwnStream path) — no duplication.
• +3 renderer tests (stop button visible when busy + click cancels; send button when idle; Esc-stop vs menu-open Esc), Agent.md renderer count 508 → 511.

Verification: renderer typecheck clean; vitest 511 passed (45 files); GUI static suite 93 pass/0 fail; doc-count guard 5/5.

[argszero]

✅ LGTM — cycle cyc20260902-205612 (1/3)

Independent review pass (authoring cycle cyc20260902-202946, this cycle differs):

• head aa031fc unchanged, CI test + test-windows PASS (run 33631587549), MERGEABLE/CLEAN, 0 prior reviews / no ❌.
• Composer busy→stop swap verified: .stop-btn CSS (red 38px circle) already existed; stop wired to injected cancelFn (default window.emrg.cancel, aligned with sendMessage/saveImage injection style).
• Esc-to-stop ordering correct: the /-menu-open Escape branch returns earlier (menu closes only, no accidental stop); the new Escape branch sits after it and gates on busyRef (live busy via ref bridge, not stale closure).
• Stop() optimistic path: clearTyping + chat.interrupted system message; daemon cancelled broadcast releases the external busy lock idempotently via existing releaseOwnStream — daemonBridge 'cancelled' case adds no system message, so no duplication.
• i18n: chat.interrupted added to both zh/en dicts (394 keys aligned, guard bumped 393→394); composer.stop pre-existed.
• Tests: +3 Composer tests (stop visible when busy + click cancels + system msg; send when idle; Esc-stop vs menu-open Esc). Local vitest 511/45 + GUI 93 + doc-count 5/5 green last cycle; Agent.md renderer count synced 508→511.

[argszero]

✅ LGTM — cycle cyc20260902-210723 (2/3)

Continuation vote (1/3 = cyc20260902-205612, distinct cycle, no ❌ between):

• head aa031fc unchanged since 1/3, CI test + test-windows still PASS (run 33631587549), MERGEABLE/CLEAN.
• No new commits, no requested changes; prior review findings stand. Re-confirmed stop() flow and Esc gating on busyRef require no revision.

[argszero]

✅ LGTM — cycle cyc20260902-212024 (3/3)

Third consecutive vote (1/3 = cyc20260902-205612, 2/3 = cyc20260902-210723, no ❌ between). Re-reviewed the diff this cycle: Composer busy state swaps send → stop button (■) wired to window.emrg.cancel, Esc stops only when busy (command menu open just closes the menu), local optimistic '⏹ 已停止响应 / Interrupted' system message, daemon cancelled broadcast idempotently releases the busy lock. +3 renderer tests, i18n both dicts (394 keys), Agent.md renderer 508 → 511. head aa031fc unchanged, CI test + test-windows PASS (run 33631587549), MERGEABLE/CLEAN. Merging.