emrg: fix GUI message display bugs — empty-state overlay, typing cursor residue, false from-other-client label

[argszero]

Fixes three GUI message display bugs (host-verified screenshots, rant 14:11) that appear after sending the first message in a session:

Bug A — welcome screen never hides

updateEmptyState() was only evaluated on session switch / new session; the streaming path (Chat.addUserMessage / delta / done) appended nodes without re-evaluating it, so #empty-state stayed overlaid on the message area.
Fix: Chat.append() and Chat.clear() now call App.updateEmptyState?.() — welcome visibility always follows chat content.

Bug B — typing cursor residue + false "from other client" label (same race)

G122 batches message_delta on a 16ms timer, but done bypasses the buffer. When the final deltas were still buffered, done reached the renderer first: groupNodes entry deleted, ownStreamRequestId cleared — then the 16ms flush delivered a stale delta that created an orphan node with isOwn=false (false label) that never receives done (cursor forever).
Fix:

1. main.js — flushDeltaBuf() runs synchronously before done / error / cancelled events (webContents.send preserves order → deltas can never arrive after the terminal event)
2. chat.js — doneRids set drops any residual stale delta (defense in depth, UUID requestIds never reused, capped at 500)
3. app.js — cancelled / stream-error clear the typing class on in-flight nodes (mid-round cancel emits no done frame, only the session-level cancelled event)

Verification

• +3 renderer smoke regression tests (welcome hides on first message / stale delta dropped / cancelled clears typing)
• npm test 91/91 green (was 88 + 3 new); Python 508 passed (incl. doc-count guard); docs synced

[argszero]

✅ LGTM — cycle (independent verification on ba8b59a)

Checked out the branch and verified each fix against the rant's root-cause analysis:

1. main.js flush ordering: flushDeltaBuf() runs synchronously before done/error/cancelled are forwarded; the pending 16ms timer is cleared first, and webContents.send preserves order on the channel — buffered deltas can no longer arrive after the terminal event. The extracted named function is also reused by the timer path (no behavior change there).
2. chat.js doneRids defense: stale deltas for finished streams are dropped before node creation (no orphan nodes, no false from-other-client label). UUID requestIds are never reused, and the 500-entry cap bounds long-running growth. clear() resets the set on session switch.
3. chat.js welcome screen: append()/clear() now call App.updateEmptyState?.(), so welcome visibility follows chat content on every add/remove — not just session switch.
4. app.js cancelled/error: clearTyping() on in-flight nodes is correct — the daemon's session-level cancelled event carries no request_id, and mid-round cancellation emits no done frame, so clearing all in-flight typing cursors is the only sound option.

Verification: npm test 91/91 on the branch (incl. the 3 new regression tests), Python 508 passed (doc-count guard satisfied, README/Agent.md/README.cn.md synced 88→91), CI run 31153709355 green.

Minor non-blocking observation: the disconnected path doesn't flush deltaBuf — a daemon death mid-stream could still deliver one buffered batch post-reconnect. Pre-existing edge, partially mitigated by the doneRids/groupNodes guards; can be addressed separately if it ever surfaces.

[argszero]

✅ LGTM — cycle (independent verification on ba8b59a)

Checked out the branch and verified each fix against the rant root-cause analysis:

1. main.js flush ordering: flushDeltaBuf() runs synchronously before done/error/cancelled are forwarded; the pending 16ms timer is cleared first, and webContents.send preserves order on the channel — buffered deltas can no longer arrive after the terminal event. The extracted named function is reused by the timer path with no behavior change.
2. chat.js doneRids defense: stale deltas for finished streams are dropped before node creation (no orphan nodes, no false from-other-client label). UUID requestIds are never reused; the 500-entry cap bounds long-running growth; clear() resets the set on session switch.
3. chat.js welcome screen: append()/clear() now call App.updateEmptyState?.(), so welcome visibility follows chat content on every add/remove — not just session switch.
4. app.js cancelled/error: clearTyping() on in-flight nodes is correct — the daemon session-level cancelled event carries no request_id, and mid-round cancellation emits no done frame, so clearing all in-flight typing cursors is the only sound option.

Verification: npm test 91/91 on the branch (incl. the 3 new regression tests), Python 508 passed (doc-count guard satisfied, README/Agent.md/README.cn.md synced 88→91), CI run 31153709355 green.

Minor non-blocking observation: the disconnected path does not flush deltaBuf — a daemon death mid-stream could still deliver one buffered batch post-reconnect. Pre-existing edge, partially mitigated by the doneRids/groupNodes guards; can be addressed separately if it ever surfaces.

[argszero]

✅ LGTM — cycle 2 (independent re-verification, different angle: test quality + edge cases)

Branch unchanged since last review (ba8b59a, +96/-13). This pass scrutinized the three new regression tests rather than the fix code:

1. Welcome-screen test asserts hidden=true after addUserMessage — valid: only the new append() hook can produce this state. (Mock limitation noted: the sandbox classList.toggle ignores the force argument, so an inverted force usage would not be caught here — consistent with the existing suite's mock fidelity, non-blocking.)
2. Stale-delta test calls Chat.handleDone directly then replays a delta — correctly isolates the renderer-side doneRids guard independent of the main.js flush; without doneRids the replayed delta would create an orphan node and fail the count assertion.
3. Cancelled test drives the real App.handleEvent path (not Chat.clearTyping directly) — verifies the full event wiring, and the querySelector-null mock fallback keeps the assertion meaningful.

Re-ran on the branch: npm test 91/91 (all 3 rant-14:11 tests green), Python 508 passed. No new commits since the previous cycle's review; CI 31153709355 remains green.

[argszero]

✅ LGTM — cycle 3 (independent re-verification, third angle: regression-surface audit)

Branch unchanged (ba8b59a). This pass audited every call site of the modified functions for regressions:

1. append() has 4 callers (user message / assistant node / system message / tool row) — all are chat content, so hiding the welcome screen on any of them is correct.
2. Chat.clear() has 4 callers (boot / rewind / switchSession / newSession); switchSession and newSession already called updateEmptyState() explicitly — the new hook makes those double calls, which is idempotent and harmless.
3. clearTyping() is wired exactly at cancelled and stream-error; the session-busy error branch correctly does NOT call it (no in-flight nodes exist on that path).
4. doneRids lifecycle is sound: add on done, check on delta, clear on session switch, 500-entry cap for long-running sessions.

Final verification on the branch: npm test 91/91, Python 508 passed, import OK, CI 31153709355 green. Three independent cycle reviews complete (fix correctness / test discriminating power / regression surface) — merge gate satisfied.