emrg: GUI main.js rewire to ConnManager — P2 per-session connections (GUI multi-session rant P2)

[argszero]

Summary

Completes P2 of the GUI multi-session rant (2026-08-10T15:07:19) connection-layer rewrite: main.js now routes through the ConnManager (per-session independent websocket connections) instead of a single global DaemonClient. Single-session behavior preserved (P2 acceptance: no regression).

Changes

conn-manager.js

• ensureDaemon() keeps a daemon-level connection (_daemonConn) for non-session commands (ping / list_sessions / set_model / github_* / ...); session connections use skipStart (never spawn)
• open(sid, projectPath, { resume }) — resume:false for brand-new sessions (daemon implicitly subscribes on first task); failed opens close the half-open connection (no leak); stale (disconnected) entries are closed + reopened instead of reused
• onOpen / onRecovered hooks — main.js attaches the sid-tagged renderer event bridge (incl. recoverAll reopen paths) and refreshes UI state after recovery
• close() marks _intentionalClose — intentional close (switch-away/delete) suppresses the renderer disconnect banner and never triggers restart detection
• Single-connection backoff (P2 acceptance): when one of N connections drops (not all), an independent backoff retry re-opens just that session; close/recoverAll cancels pending retries

daemon_client.js

• G65 own-stream lock moved into each DaemonClient instance (ownStream / ownStreamRequestId set by sendTask(stream:true), released on own done / timeout / session-busy error / own cancelled / disconnect) — multi-session independent locks

main.js

• Removed global client / ownStream / ownStreamRequestId / deltaBuf state; all IPC handlers route through activeConn() / requireConn() / openSession()
• emrg:sendMessage auto-opens the session connection (resume:false) on first message
• emrg:switchSession opens via connManager (G106 passive-delete recovery kept), closes the previous session's connection (single-session: one active conn)
• emrg:cancel targets the active connection and releases its lock
• Restart recovery: session drops → connManager recoverAll → onRecovered → sessions/pong/status refresh; daemon-level retry backoff kept for the no-session-open case (G41 reworked)
• Events forwarded to renderer with sid (backward-compatible; renderer ignores extra field)

Tests

• daemon_client 37→43 (+6 ownStream: set-on-send, own-done release vs broadcast-done hold, timeout release, session-busy release, cancelled release, disconnect release)
• conn-manager 8→17 (+9: ensureDaemon reuse, open with resume:false, failed-open closes conn, onOpen hook incl. recoverAll, onRecovered hook, intentional-close no-restart, stale-conn reopen, single-conn backoff)
• GUI 127→142 (43 dc + 17 cm + 22 + 32 + 15 + 7 + 3 + 3); 680 pytest green (doc-count guard); node --check clean

[argszero]

✅ LGTM — cycle 20260810-162905

Verified locally on the PR head: GUI suite 142/142 (43 daemon_client + 17 conn-manager: per-conn ownStream set/release incl. broadcast-done negative; ensureDaemon daemon-conn reuse; resume:false for new sessions; failed-open conn cleanup; onOpen/onRecovered hooks; intentional-close suppresses disconnect banner + never triggers restart detection; stale-conn close+reopen; single-conn backoff re-opens only the dropped session), 680 pytest green (doc-count guard reads 142 breakdown), node --check clean. Design matches the rant P2: connManager = daemon lifecycle owner, per-session connections, delta batching per connection (#626), restart recovery (#627), single-drop independent backoff.

[pm25coder]

Tested this PR on Windows (local checkout of feature/gui-p2-main-rewire @ c14e7a8):

[argszero]

✅ LGTM — cycle 20260810-165036

Re-verified on head c14e7a8 (master unchanged since): GUI suite 142/142 (43 daemon_client + 17 conn-manager — per-conn ownStream set/release incl. broadcast-done negative, ensureDaemon daemon-conn reuse, resume:false new-session path, failed-open conn cleanup, onOpen/onRecovered hooks, intentional-close suppresses banner + never triggers restart detection, stale-conn close+reopen, single-conn backoff re-opens only the dropped session), 680 pytest green (doc-count guard 142 breakdown), node --check clean. PR CI 31371726743 SUCCESS.

[argszero]

✅ LGTM — cycle 20260810-164959

Verified independently on the PR head (c14e7a8): GUI suite 142/142 (43 daemon_client: per-conn ownStream set on stream send, own-done release vs broadcast-done negative, timeout release, session-busy release, cancelled release, disconnect release; 17 conn-manager: ensureDaemon daemon-conn reuse, resume:false no-resume, failed-open conn cleanup, onOpen incl. recoverAll reopen path, onRecovered refresh, intentional-close suppresses banner + never triggers restart detection, stale-conn close+reopen, single-conn backoff reopens only the dropped session), 680 pytest green, doc-count guard passes (142 breakdown synced in Agent.md/README/README.cn), node --check clean.

Review notes: main.js now fully routes through ConnManager (activeConn/requireConn/openSession) with per-session ws connections; daemon-level _daemonConn serves non-session commands; G65 own-stream lock is per-connection. Also noticed scheduleReconnect now applies the backoff delay (was hardcoded 1000 — latent bug fixed). Design matches the GUI multi-session rant P2: connManager = daemon lifecycle owner.

[pm25coder]

Post-merge verification on master 2840017 (fast-forward from 4f70a79):