Roster: wire the 'poke' button (wake / continue-loop nudge)

Broader finding (2026-07-02, while routing feedback #49 to orchestrator per operator): the feedback→orchestrator ROUTING is broken the same way as poke/wake. `flower:feedback-promote --to=orchestrator` fails at `solo-cli kv get routed_pending failed: unknown command: kv` — the routed_pending queue is stored in Solo KV and the wake uses timerSet, and solo-cli 0.9.3 exposes NEITHER `kv` NOR `timers`. So the entire feedback→orchestrator coordination path (routed_pending queue + orchestrator wake) is non-functional on the current CLI, not just timerSet (bug #49). The fix CONVERGES with the orchestrator-mediated mechanism this brief needs: move the intent queue (routed_pending, poke_requests) OFF Solo KV into a flower DB table the app owns, and have the orchestrator drain it on its loop + do the MCP-level wake. Recommend one shared "orchestrator-mediated coordination" substrate: a DB-backed intent/queue table + orchestrator drainer, serving poke (#71), feedback routing (#49 / FeedbackPromotionService), and self-reset (#76). NOTE: the operator-hit todo-title>200 bug was a separate, now-fixed issue (commit 569b372) — this kv/timers break is the next layer.

Feasibility gate (Codex 994 + orchestrator verification): the direct app→solo-cli poke is NOT possible. solo-cli 0.9.3 exposes projects/agents/processes/todos/scratchpads but NO `timers` or `send-input` — the app cannot inject a wake turn into a Solo process; only an MCP-connected agent can (mcp__solo__timer_set / send_input, which is how the heartbeat wakes daemons). `SoloClient::timerSet()` is dead (calls non-existent `solo-cli timers set`); its live caller feedback_promote's wake is broken (filed as a bug). RE-SCOPE 66b to an ORCHESTRATOR-MEDIATED poke: the /roster button writes a durable poke intent (a poke_request row, or reuse the routed_pending queue pattern FeedbackPromotionService already writes); the orchestrator daemon drains the queue on its loop and performs the actual MCP injection (timer_set/send_input) against the target daemon's solo_process_id. Latency = up to one orchestrator heartbeat (~13m) rather than instant — acceptable for a nudge. This mechanism is SHARED with self-reset (#76) and repairs the feedback_promote wake. NOTE: spawn (66c) is app-doable (solo-cli `processes spawn` exists); wind-down (66d) is flag-based (daemon polls) so also app-doable — poke is the one that specifically needs injection. Status → refining pending operator decision on the mediated mechanism.

Child of #66, maps note #6. Poke = manually fire the same wake/continue-loop nudge the idle heartbeat timer sends to a daemon (inject a fresh user turn into its Solo process / timer_fire), so the operator can nudge a parked or idle daemon to resume its loop on demand. Confirmed semantics: NOT a checkin-only. Wire the existing poke button on /roster to this action via the Solo bridge (same path spawn uses) or a small daemon-nudge service. S, near-trivial once the send-input/timer path is chosen.