Durable stateful backend: cold re-attach, remove pi-mirroring, auto-mirror fan-out (#26)

* feat(server): durable stateful backend — cold re-attach, drop mirroring, auto-mirror fan-out

Make the makit server a durable, multi-device WS backend where any
device sees every session's activity live, and sessions survive a
server restart.

## Cold-session re-attach (resume the live agent after a restart)
- Persist `resumeSessionPath` alongside session metadata (new nullable
  SQLite column + `SessionMeta` field; back-compat ALTER for old rows).
- `SessionManager.reattachSession(id)`: find the cold (rehydrated)
  session, rebuild the real adapter (pi resumes from its on-disk
  transcript via `--session`), swap it in with `replaceAdapter`, and
  start it. Non-pi / no-resume-path sessions stay history-only and keep
  emitting the clear `session.error`.
- The durable `seq` space is preserved: new events continue after the
  persisted max (the store assigns the next seq), so mobile↔desktop
  handoff survives a restart end-to-end.
- Wired via the `session.attach` cmd: `{sessionId}` re-attaches a cold
  session; legacy `{projectId, piSessionId}` still attaches an on-disk
  transcript (wire back-compat).

## Remove pi-mirroring (redundant now that makit owns/re-attaches pi)
- Delete World B (file-tailing MirrorAdapter, `makit mirror`,
  `makit attach --pane`) and World D (`makit-mirror` extension, host
  sessions, IngestAdapter), plus the `pane/` + `mux/` infra and the
  standalone slash-palette fetcher (`commands.ts`).
- `spawnSession` now always uses the headless PiAdapter; host token /
  `~/.makit/host.json` / auto-install wiring removed.

## Auto-mirror fan-out (every device sees every session)
- `SubscriptionHub.fanout` delivers each session event to ALL authed
  clients, no explicit `sub` required. The `subscribed` set now only
  governs history replay-on-`sub` (with fromSeq cursor) and
  AskUserQuestion/ReverseRpc prompt routing.

## Tests
- reattach: persist a pi-backed session + events → new manager on same
  store → reattachSession rebuilds a live adapter → next event gets
  seq = max+1 (continuity), history intact.
- auto-mirror unit + a two-client integration test against a live
  startWsServer proving an event from one client reaches another that
  never subscribed.

pnpm typecheck clean; pnpm test 228 pass / 0 fail.

* ci: gate expensive macOS e2e to workflow_dispatch only

macOS runners cost far more than the Linux gates, and the cheap Linux
checks (server-ci unit tests + typecheck, protocol-contract, and the
headless real-pi-pinned smoke) already cover regressions on every PR.

- integration-ci.yml (stub-e2e, macos-latest): drop the pull_request /
  push triggers — the workflow is now workflow_dispatch only. (A manual
  dispatch already ran both stub-e2e and stub-e2e-coverage, so dispatch
  behaviour is unchanged; only the per-PR/push runs go away.)
- real-pi-e2e.yml (real-pi-app, macos-15): gate to
  `if: github.event_name == 'workflow_dispatch'`. The Linux real-pi-pinned
  (PR/push blocking gate) and real-pi-canary (weekly schedule) are untouched.

Run them on demand from the Actions tab (or `gh workflow run`) when a
change warrants a full-stack check.

* chore: re-trigger CI after CLA fix (empty; drops on squash-merge)

Author: leduckhc

.github/workflows/integration-ci.yml

@@ -3,23 +3,12 @@ name: Integration E2E (stub adapter)
 # Keyless full-stack integration: real Flutter app ↔ real TLS WS server backed
 # by the StubAdapter (server/src/adapters/stub.ts) on an iOS simulator. No LLM,
 # no API key — the stub replies with scripted scenarios (echo, STREAM, THINK,
-# MARKDOWN, ASK_QUESTION, ASK_MULTI), so this runs on every PR/push as a
-# blocking gate. The real-pi path (needs a live pi + API key) stays in
-# real-pi-e2e.yml, opt-in only.
+# MARKDOWN, ASK_QUESTION, ASK_MULTI). macOS runners are expensive, so this is
+# MANUAL (workflow_dispatch) only — not a per-PR gate. The cheap Linux gates
+# (server-ci unit tests + typecheck, protocol-contract, real-pi-pinned headless
+# smoke) cover regressions on every PR. The real-pi path stays in real-pi-e2e.yml.
 
 on:
-  pull_request:
-    branches: [ main ]
-    paths:
-      - 'app/**'
-      - 'server/**'
-      - '.github/workflows/integration-ci.yml'
-  push:
-    branches: [ main ]
-    paths:
-      - 'app/**'
-      - 'server/**'
-      - '.github/workflows/integration-ci.yml'
   workflow_dispatch: {}
 
 jobs:

.github/workflows/real-pi-e2e.yml

@@ -7,7 +7,8 @@ name: Real-pi E2E (fake model)
 #   real-pi-pinned       — pinned pi version, headless smoke on Linux, blocking on PRs.
 #                          Failure = our bug.
 #   real-pi-app          — real pi + Flutter e2e on iOS simulator (macos-15).
-#                          Full stack proof. Pinned pi, PRs only (slower/costlier).
+#                          Full stack proof. Pinned pi, MANUAL (workflow_dispatch)
+#                          only — macOS runners are expensive, so not a per-PR gate.
 #   real-pi-canary       — pi@latest on a schedule. Failure = a pi upgrade broke the
 #                          integration; opens an issue so we can bump the pin.
 
@@ -81,9 +82,9 @@ jobs:
 
   # Full-stack on macOS: real pi + real Flutter e2e via `tool/e2e.sh --mode real`
   # on an iOS simulator. Proves the complete makit↔pi integration on the target.
-  # Pinned pi, PRs only (slower, costlier).
+  # macOS runners are expensive — MANUAL (workflow_dispatch) only, not a per-PR gate.
   real-pi-app:
-    if: github.event_name == 'push' || github.event_name == 'pull_request'
+    if: github.event_name == 'workflow_dispatch'
     runs-on: macos-15
     timeout-minutes: 40
     steps: