feat(suggest): on-demand next-prompt suggestion deck (orb → fan → stack)

[edwin-zvs]

What

An on-demand "suggestion deck" for any harness session (spec 0109): when a session is awaiting input, a corner orb appears on the session view. Click it (or C-x s) and the daemon generates a hand of suggested next prompts — one top pick plus 3–4 generated verb directions, each holding concrete prompt cards. The hand renders as a quarter-arc verb fan from the corner; picking a verb opens its card stack; accepting anything sends the text through the ordinary session.input path.

Design decisions (see specs/0109-suggestion-deck.md)

• On-demand only — generation runs on orb activation, never automatically at turn end. Idle sessions cost nothing. [suggest] enabled = false disables it entirely.
• Same-harness generation. Smith sessions use a new --suggest-mode process one-shot (mirrors --title-mode). Natively-forking harnesses (claude / codex / opencode / grok / pi) generate through a hidden fork probe of the target session: the harness resumes its own native conversation, so the provider's prompt cache already covers the full history — the request is one appended message, and the model predicts from complete context instead of a rendered tail. Other harnesses get a fresh hidden probe fed a rendered transcript tail. Probes are SessionKind::UsageProbe (hidden) and deleted after use.
• Ephemeral event. SessionEvent::Suggestions is broadcast-only, never persisted (same treatment as AgentStatus). Any new turn / user message / terminal state invalidates the hand, and a monotonic per-session generation counter discards stale in-flight results.
• Typing always wins. The overlay only consumes keys while open; any printable key closes it and takes its normal route to the PTY.
• One shared contract. SuggestionHand::parse_loose (protocol crate) does the loose-in/strict-out parse + clamping for every generator.

Surface

• protocol: SuggestionHand/Verb/Card, SessionEvent::Suggestions, session.suggest IPC, shared prompt + parser
• daemon: request_suggestions + smith one-shot / fork-probe generation, [suggest] config
• cli (TUI): orb badge (idle → spinner → count), fan, stack, keyboard (C-x s, arrows, ⏎, ⌫, esc) + full mouse support, ToggleSuggestions key action
• smith: --suggest-mode

Demo (real end-to-end: smith session → orb → fan → stack → send → new turn)

▶ demo video

Verb fan (labels fully generated from the session):

Card stack for the "recount" verb:

(Media is pinned at commit 679012b, removed from the branch tip in the next commit.)

Testing

• cargo test --workspace: green (1070 cli + daemon + e2e all pass)
• Live smoke against an isolated daemon: session.suggest acks started:true, hidden probe never appears in construct list, and a real smith session produced a structurally-valid hand (top + 4 verbs × 2 cards) that broadcast to a subscribed client
• vhs recording above exercises the full TUI loop including accept-and-send

Follow-ups (not in this PR)

• Web UI rendering of the hand (it already receives the event via subscribe.events)
• Fill-in-the-blank template cards; acceptance-feedback context
• Anthropic cache_control breakpoints in the smith one-shot path

🤖 Generated with Claude Code

[edwin-zvs]

Fixed two issues found trying it live on the TUI:

1. Mouse clicks on the orb now work over mouse-grabbing children. Clicks over the pane were piped into the child PTY (e.g. Claude Code fullscreen) before construct's hit-testing ran, so the orb never responded to the mouse. The deck now takes priority ahead of forward_mouse_to_child, same as the session-title menu and URL clicks. While the overlay is open, the whole pane acts as its backdrop (click-away closes instead of leaking into the child).
2. The verb fan is now a true quarter-circle around the orb. The old formula pinned chips to fixed rows with only a few columns of bend — it rendered as a near-vertical column hugging the right edge. Chips now sweep a 15°–85° arc with aspect-corrected radii: top pick left of the orb, last verb above it, one chip per row, compact stair fallback for short panes.

Updated fan (v2 video):

[edwin-zvs]

The orb is now a shaded ASCII ball with a face — software-rendered in the ASCII-shader style: per-cell diffuse + specular lighting from an orbiting light, luminance mapped onto a ·:-=+*#@ ramp over a phosphor gradient of the theme accent, transparent outside its silhouette. It blinks/winks at rest, spins moon-phase eyes while thinking (the light orbits faster too), gets sparkle eyes when a hand is dealt. Small panes fall back to the one-row kaomoji. Also in this batch: fan Up/Down now match the visual direction.

resting / thinking / fan open (v3 video):