Pick attach vs resume from actual session liveness when jumping into a session

[MJohnson459]

a on a review task whose claude --bg session was still resident failed with
the supervisor's refusal (Session <uuid> is currently running as a background agent), because jump_into_session read its verb from task state alone:
Running → attach, Review/Stalled → resume. Session liveness does not track
task state in either direction — a bg session commonly outlives running (the
operator jumps into a review task's session to ask for a rebase, DESIGN.md §8),
and a session can die while its task is still running, where attach has
nothing to join.

What changed

• New crates/voro/src/session_probe.rs holds the run-and-parse of an
  agent's sessions verb, previously duplicated in reconcile.rs and
  dispatch.rs. run_sessions_command(cmd, cwd) and listing_says_live
  serve reconcile (which keeps its own per-agent listing cache) and dispatch's
  ref capture (which keeps its cwd/startedAt filtering); session_is_live
  serves a caller holding no listing. Liveness stays three-valued: None when
  there is no sessions verb, no captured ref, or the listing would not run.
• jump_into_session probes on keypress — synchronous, since the TUI is
  about to hand the terminal over anyway — and picks the verb from the answer,
  falling back to the task-state mapping only when liveness is unknowable.
• Verb availability is a further fallback: jump_verb takes whichever
  template the agent defines when its preferred one is absent (the built-in
  codex has only resume), and errors naming both verbs only when it
  defines neither.
• The existing gates are untouched: jump-in is still offered only on
  running/review/stalled, and the no-session / no-ref status lines are as they
  were.

No DESIGN.md change: §8 already describes the operator attaching to a review
task's still-open session, which is what this makes possible.

Verification

cargo test --workspace and cargo clippy --workspace --all-targets -- -D warnings pass. The verb choice is a pure function with unit tests over the
liveness × state × available-verbs grid; the app-level jump-in tests now drive
a against a canned listing in each configuration — review + live → attach,
running + finished → resume, no sessions verb → state mapping unchanged,
one-verb agent → that verb, no verbs → both named in the message.

Also driven through the real TUI against a scratch database with a stub agent:
on a review task whose listing showed the session blocked, a logged
attach: echo I-AM-ATTACHING '<ref>' in launches.log; rewriting the listing
to state: done and pressing a again logged the resume command.