Design: a contribution has no outer loop, and its arming comment answers its first gate

Phase 2 of 4 (bugfix → design → testing plan → tasks). Derives from the approved bugfix.md. The human gate for this work item is the pull request.

Overview

Two small, independent changes, each in the module that owns the fact it is about.

Requirement	Mechanism
R1.1–R1.4	graph/hooks/selection.py: one predicate, _asks_surface(ctx), read at the three places the surface question exists — render, parse, confirm — with the frozen record carrying ""
R1.5	graphlink.py: GraphContext.loop, and one branch in render_graph_context
R2.1–R2.5	graphlink.py::GraphLink.on_spawn: evaluate a human start node once, with the spawning event; dispatcher.py::_spawn_tmux passes routed

Nothing changes about which edges exist, who may answer a gate, or what a hook may do.

flowchart LR
    subgraph before["before"]
        A1["the-loop contribute<br/>(+ goal)"] --> B1["spawn → start()"]
        B1 --> C1["goal-definition<br/>(never evaluated)"]
        C1 -.->|"some later event"| D1["phase-selection<br/>+ surface row"]
    end
    subgraph after["after"]
        A2["the-loop contribute<br/>(+ goal)"] --> B2["spawn → start()"]
        B2 --> C2["human start node?<br/>→ advance(event)"]
        C2 --> D2["phase-selection<br/>no surface row"]
    end

D1 — the surface question belongs to loops that have an outer loop

_asks_surface(ctx) is ctx.graph.name != PDLC_CONTRIBUTION_LOOP. Three properties made this the right shape:

• Derived from the compiled graph in context, the same source _phase_rows reads. The checklist a user sees, the vocabulary their reply is validated against, and the record that is frozen therefore cannot disagree about which loop this is.
• Named negatively — "not the contribution loop" — rather than positively. A repository may supply its own pdlc-work-item-loop.yaml; it owns a work item, it has an outer loop, and it keeps the question. Only the one graph a repository cannot override loses it, which is also the only graph we can be certain has no outer loop.
• pdlc-pr-loop needs no case. It has no phase-selection node and never reaches this hook; the existing docstring already says so, and this change does not make that claim load-bearing.

The absent row is replaced by a sentence, not by silence: a reader of somebody else's ticket should not have to work out where this conversation is going to happen.

The record carries "", not the default. NO_SURFACE = "" distinguishes never asked from offered and left alone, and it costs nothing downstream because Runtime._record_selected_skips already writes state.surface only for a truthy value — so graph-state.json keeps its pre-issue-199 shape, and no reader learns a third literal. _confirmation gains the same guard: claiming "the outer loop happens here, on this work item (the default)" would report a choice the human was never offered.

D2 — the prompt does not name an outer loop for a contribution

GraphContext gains loop, read from the same graph-state.json the rest of the context comes from, and render_graph_context branches on it. This is the reader's half of D1: the gate stops asking, and the prompt stops instructing. Both go through _is_contribution(loop) rather than an inline comparison — two very different callers asking one question.

D3 — a human start node is evaluated once, at spawn

def call(rt, item):
    report = rt.start(item, work_item.ref)
    self._bind_session(rt, item, session_id, runner)
    if report is None or not self._entered_a_human_gate(rt, report):
        return
    rt.advance(item, ref=work_item.ref, event={"comments": comments_from(routed)})

Four deliberate narrowings, each answering "what could this break?":

1. report is None returns. Runtime.start is pointer-idempotent, so this is exactly "a fresh entry happened". A respawn, a redelivered spawn and a session restarted after a crash all evaluate nothing (R2.3).
2. Human nodes only. An agent node's exit chain gates checked-in artifacts; running it before the session has done any work would count a failed attempt against work that has not begun, and max_attempts escalation is a real consequence (R2.4).
3. The event rides along. comments_from(routed) is the same translation on_event uses; routed=None yields []. classify-goal re-reads the thread anyway, so this is not what makes the fix work — it is what keeps the two paths from disagreeing about what a gate is handed.
4. Inside _guarded, and therefore inside the state lock and the blanket except. The advance runs after _bind_session, so a session: inherit gate resolves against the session that was just spawned. Nothing here re-acquires the lock (Runtime.advance does not; only Runtime.complete does).

_entered_a_human_gate fails closed on anything unreadable: not evaluating leaves the work item where the pre-issue-199 code left it.

Why not fix it in the dispatcher instead (forward the arming comment after _apply_control)? Because that would change what a control command is: the whole point of issue-106's boundary is that a keyword is executed by the-loop and not handed to the harness. Delivering it as an event would also render a second prompt into the freshly spawned session. The graph coupling is the right place: it already owns "an event reached this work item, let the graph see it".

Why the spawn prompt is not re-rendered. _spawn_for resolves context before the spawn by design (issue-148 D5: reads before, writes after), so a prompt whose gate resolved milliseconds later is momentarily stale. It is not corrected here: the session's own the-loop check/graph complete read live state, the stale line is a status echo rather than an instruction, and re-ordering the spawn around the graph would trade a cosmetic staleness for the possibility of entering a graph for a session that never started.

Trade-offs considered

Option	Why not
Give pdlc-contribution-loop its own selection hooks	Two copies of a gate that must stay identical in authorization, freezing and provenance — the failure mode is silent divergence
A surface: none marker in the graph YAML	A repository cannot override this graph, so the marker would be configuration nobody can set; the loop's identity already carries the fact
Record surface: "work-item" for contributions	Reads back as a choice that was made; the whole complaint is a question that should not have been asked
Evaluate every start node at spawn	Breaks R2.4 — agent nodes would block/escalate on artifacts nobody has been asked for yet
Have classify-goal poll, or the poller re-check	Adds a scheduled path for something the event that already exists can answer