Detect partner dropout and exit the survivor with a payment path (C&WG)

[htsukamoto5]

Closes #5.

Stacked on #16 (feat/datapipe-egress-and-identifiers), which is stacked on #14. It needs #16's Pipeline.flush() and #14's submissionBlockHTML / wireSubmissionButton, so review order is #14 → #16 → this. The diff here is only the dropout work.

The problem

The survivor of a dropout did not hang — they ground, which is worse, because it burns their time before they give up. round_timeout arms at round start, so when it fires the round ends and the timeline advances to the next one with a fresh timer. Five more rounds at 180s is ~15 minutes in front of a fully interactive board that never responds, with no copy anywhere telling them what happened.

And then, as the #14 review pointed out, it got worse: they reached "Game complete!", were told they scored 0 / 0 across 0 trials, and were handed the complete code. One dropout burnt two payments, miscoded the reconciliation, and generated a support ticket.

Detection: silence, not timeouts

round_timeout alone is not the signal. It is an unconditional wall-clock bound (corrected in c2c69c2 on the base branch), so it fires identically for a partner who left and a pair who are simply slow — and trial 1 of a 12-figure board is the longest in the study. Counting timeouts would abort real dyads at exactly the point the effect is largest.

So a round is silent only if it timed out and the partner sent nothing during it. The plugin records message_count (every message that round) and messages_sent (mine) on a per-round chat channel, so the difference is the partner's activity in that round alone. Verified against the vendored bundle rather than assumed:

message_count: g.length,
messages_sent: g.filter(m => m.senderId === b).length

CONFIG.DROPOUT_SILENT_ROUNDS: 2 — suggested, not settled. Requiring partner silence is what makes 2 safe rather than reckless: a slow dyad that is still talking is never aborted, so the threshold does not have to buy its safety with more dead minutes. At 180s that releases a survivor after ~6 minutes instead of 15.

The costs are asymmetric and worth stating explicitly: a false positive ends a live dyad and pays two partial codes; a false negative leaves a survivor grinding. The tests are weighted toward the false-positive side, because that is the direction that silently destroys good data.

On trip

1. Record the outcome and the round it fired on.
2. jsPsych.abortCurrentTimeline(). gameLoop's own conditional_function is evaluated once, before the first round, so it cannot end the loop mid-game — this is what actually stops the remaining rounds.
3. Show an honest screen: what happened, that it was not their fault, that payment is unaffected, how many rounds they completed, and the partner_dropped code.
4. Flush data. The submit button is wired before the flush resolves, so the redirect never waits on OSF.

Also fixes an existing miscoding

doneTrial ran unconditionally, so a spectator — who had just been told the game was full — fell through to "Game complete!" reporting 0 / 0 and received the complete code. That predates this PR; the survivor path just made it visible. Both terminal screens are now conditional and mutually exclusive, so exactly one runs and each carries its own code.

Data

ended_reason, n_trials_completed, n_trials_scheduled, dropout_detected_at_round on every row. n_trials_completed counts only ended_by: "submit" — a timed-out round has a null assignment and is not a trial. This is what lets a partial dyad be filtered rather than merely present (#8), and it records how a session ended in exactly the partial data most likely to be all we have.

Testing

node tests/dropout.test.mjs — 16 checks, no dependencies. The detector is extracted from the experiment file and driven with synthetic round data, so it is tested as written rather than as re-implemented in the test.

Never aborts: a clean 6-round game; a slow-but-talking dyad across three timed-out rounds; a single silent round; non-consecutive silences; a partner who speaks again after a silence; a partner talking while I am silent. Does abort: two consecutive silent rounds, mid-game dropout after real trials, and exactly once rather than per subsequent round. Plus the configurable threshold and degenerate data with the chat fields missing entirely.

node tests/pipeline.test.mjs still passes 39/39.

Not in here

• Pairing: lobby waits unbounded — add timeout, no-match exit, and payment path #6 — the lobby timeout and no-match exit. It reuses this screen's shape, so it is cheap now that this exists.
• Real disconnect detection. This is deliberately the no-presence-API version; Verify whether adapter-multiplayer-firebase exposes presence / onDisconnect #11 upgrades it to a fallback once a presence signal exists, and should not block it.
• The partner_dropped code is still an empty placeholder until Create the DataPipe experiment and a draft Prolific study — critical path #15, so this path warns at startup and shows the configuration-error screen rather than a code.

🤖 Generated with Claude Code