Make xmd repl operable from its visible terminal interface #870

Story

As a person opening xmd repl for the first time, I want the terminal to show
what I can do, what I am currently reading, and why an action is unavailable,
so I can run entries, follow Agent work, answer interactions, inspect history,
and return to the live execution without memorizing an undocumented key map.

Common path

Run xmd repl in an ordinary terminal. Starting from the empty screen, use only
what the screen itself communicates to:

1. type and submit Entry 1;
2. recognize whether it is running, waiting, failed, cancelled, or complete;
3. move between Entries and Sessions and return to the same selected entry;
4. open and complete an Elicit, Plan review, or Agent permission when one is
   waiting;
5. inspect an earlier History position and return to [live]; and
6. submit Entry 2 after Entry 1 has completely settled.

The same path remains understandable and operable in a normal wide terminal and
at the supported 72x20 minimum. A mouse may make an action faster, but the
keyboard path is complete.

Current gap

The delivered REPL has the execution, reconstruction, Entries, Sessions,
History, forms, permissions, focus, scrolling, and responsive-layout machinery
needed for this journey. Its evidence generally drives known control keys or
preselected actions, however. No production Story owns the first-use contract:
that an unfamiliar person can discover the interaction model from the rendered
terminal and finish the common path without reading the specification or test
helpers first.

That distinction matters because a control can be mounted, targetable, and
correct once activated while the product remains unusable to someone who cannot
tell how to reach it, what it will do, which state is selected, or why pressing
Enter changed nothing.

Contract

The screen teaches the interaction

• The visible interface communicates the keyboard operations needed for the
  common path: edit, move focus forward and backward, activate, dismiss, and
  leave the REPL safely.
• A person does not need a hidden shortcut to reach or activate any essential
  control. Pointer activation remains equivalent to activating the same visible
  control from the keyboard.
• Guidance stays concise enough to leave the transcript, contextual drawer, and
  fixed History footer usable at 72x20. If guidance has a detailed form, the
  way to open and close it is itself visible.
• Instructions describe the action available in the current context. They do
  not advertise a control that is absent, disabled, historical, or outside the
  mounted surface.

The screen explains its state

• The active surface, selected entry, focused control, live-versus-historical
  position, and entry outcome are distinguishable in the rendered frame rather
  than only in the URL or application model.
• When Run is unavailable because an entry is active, teardown is unfinished,
  or History is being inspected, the refusal remains visible and the draft
  remains byte-for-byte intact.
• Waiting questions and permissions are discoverable without opening
  themselves or stealing focus. After activation, the drawer identifies what
  it belongs to, exposes every required field or choice, and communicates how
  to submit or dismiss it.
• Windowed Entries, Sessions, forms, permissions, and History make both the
  existence and direction of additional content apparent. Every essential row
  remains reachable after resize, and the first scroll action moves the visible
  window.
• Returning from Sessions, a drawer, or History restores a meaningful visible
  locus instead of leaving focus on an absent control.

Existing truth and ownership remain unchanged

• The Journal and canonical location remain the only durable execution and view
  state. Help text, visual emphasis, and other presentation state do not become
  durable records or change replay.
• Background Agent activity, permission arrival, entry completion, and terminal
  resize may repaint status but do not select themselves, change the Sessions
  filter, open a drawer, move the History position, or steal focus.
• Narrow composition still mounts only the routed outlet. Making navigation
  discoverable does not restore hidden panes or make clipped controls
  targetable.
• One command owns one terminal reader and at most one live entry task. This
  Story changes how the existing state is understood and operated, not its
  execution or persistence semantics.

Acceptance

• A fresh real-terminal journey, given only the command to run, completes the
  common path using instructions and controls visible in the terminal.
• The journey proves forward and backward focus traversal, Enter activation,
  Escape dismissal, and pointer activation against the exact rendered frame.
• At both a normal wide size and 72x20, the frame exposes a route between
  Entries and Sessions, access to History and [live], access to every waiting
  interaction, and a way to leave the REPL without obscuring the selected
  transcript or draft.
• A multi-entry Plan journey distinguishes queued, streaming, waiting,
  terminal, recorded, failed, and completed states where they occur. Selecting
  an earlier entry hides the live transcript overlay but never removes the
  execution-wide Sessions chronology.
• Refused Run, invalid form submission, unavailable historical action, and a
  too-small terminal each explain the problem without dropping the draft,
  mutating retained history, or leaving an invisible actionable control.
• A cold reopen renders the same selected state and usable controls without
  contacting a provider. Presentation-only additions do not change the Journal
  bytes.
• Removing the visible interaction guidance or selected/live-state indication
  makes the first-use journey fail even though the underlying actions still
  exist. Mounting an essential control outside the placed frame likewise fails
  its rendered-frame evidence.

Evidence

Drive the production command boundary, terminal input decoder, mounted Freedom
tree, layout, renderer, real packaged Plan, Agent stream, contextual drawers,
Journal, and canonical location. Assert what the terminal actually renders and
what a pointer can target; reducer descriptions or action keys alone are not
sufficient evidence for discoverability.

Keep asynchronous waits bounded by real-time deadlines and report the rendered
screen on timeout. Preserve the existing wide, medium, narrow, cold-reopen, and
terminal-teardown evidence while adding the first-use journey.

Relationships

• Child of Quest Quest: Run, inspect, and resume XMD execution in the REPL #827.
• Begins after PR ✨ Slices C and D of #827: a draft, an Entries catalog, and the journey that proves them #869 and authoritative CI for its merge commit are green.
• Builds on the delivered one-entry kernel (Run and reconstruct one XMD entry in the REPL #848), Agent and Plan journey
  (Run a Plan and follow its Agent sessions in the REPL #854), concurrent document work (Run spawned document work concurrently with <All> #855), and sequential-entry catalog
  delivered through PRs ✨ Slice A of #827: project sequential REPL entries from one Journal #866–✨ Slices C and D of #827: a draft, an Entries catalog, and the journey that proves them #869.
• The Product Owner's XMD REPL Terminal Interface study remains the visual and
  interaction reference; the implementation is the production terminal, not a
  new mockup.

Out of scope

• Forking from a historical position or defining the durable semantics of a
  fork.
• Renaming, deleting, importing, or otherwise managing Agent sessions.
• Provider-native terminal handoff, provider-TUI mirroring, or xmd tail.
• A new Journal record family, snapshot, sidecar state, public terminal UI
  package, or replacement renderer.
• Adding the interactive REPL to Workflow.

Reference resources

• XMD REPL Terminal Interface corrected.zip is the canonical Product Owner animation and keyboard-study bundle. SHA-256: 86be64318121546f3208342edaf17f0a3aed35f3fde2766b523423e6c708c77b.
• Issue Render the REPL interaction study in a real terminal #838 and closed experimental PR 🧪 POC stack: render, route, compose, pause, and reconstruct the XMD REPL (#838–#842) #844 record the real-terminal adaptations and findings derived from that reference. Their disposable implementation is evidence, not a production starting point.