feat(vigil): interactive TUI wiring (slice 3 of 5)

[allen-munsch]

Slice 3 of 5 in the vigil decomposition (stack #812, stacked on vigil/02-core-runtime).

Wires the slice-2 runtime into the interactive TUI loop behind the opt-in vigil feature. No dependency changes; Cargo.lock is untouched.

What this does

• run_interactive gains the four vigil receivers (wake / observance / ctl / hook) and drains plugin hook dispatch requests every iteration, so on-vigil-event / on-vigil-reap hooks fire even when a rite fails, a vigil is paused, or a dispatch happens from commands mode.
• A vigil-wake select! arm observes a reap, stores pending-observance metadata, and launches the agent turn; the post-turn handler (VigilBits) then dispatches on-vigil-observance with :response and :exit and releases the in-flight flag.
• decide_post_done_action gains a VigilSleep outcome so an active vigil suppresses loop/followup auto-restart. This is cfg-gated and is a no-op when the feature is off.
• Left panel: PanelMode::Vigil, VigilStatusRow, VigilLeftPanel widget, Scene.left_panel_mode + vigil_data, /panel vigils, and /vigil subcommands (add / start / stop / status / rest / pause / resume / remove), with live status polling through VigilCtl::StatusReq.
• Idle poll slows from 20Hz to 1Hz while a vigil is active.

Testing

• Unit tests: test_post_done_action (including VigilSleep ordering), panel_data::duration_short_*, and the full scene::tests set all pass.
• Local CI reproduction: cargo fmt --check, the clippy matrix (--all-targets -- -D warnings for default, windows-default, sandbox-microvm, all-features), and the build matrix (10 configs with RUSTFLAGS="-D warnings") are all green. Feature-OFF (windows-default / no-plugin) stays warning-clean.
• cargo test --bin dirge --all-features: 6041 pass; 13 failures are all pre-existing sandbox::microvm / e2e_microvm_attach container-build tests, unrelated to this slice.

Known gaps

• on-vigil-observance dispatch assumes the plugin feature is present when vigil is; without it the hook is silently skipped (same as the slice-2 headless path).
• The status-poll oneshot round-trip is awaited at the top of each UI iteration; under a slow keeper this could add latency, though the keeper answers StatusReq synchronously.
• No interactive TUI smoke test covers the full wake→turn→observe cycle yet; that lands with the slice-5 e2e fixtures.