Triplet view collapses requests without valid prompt token IDs #572

What happens

The format=triplet event view currently groups every model_request by prompt_token_ids and keeps the last event for each key. When those IDs are absent or empty, every request gets the same empty key, so unrelated calls in one rollout are silently collapsed.

Malformed values can be worse: a nested list reaches tuple(prompt_token_ids) and raises TypeError, turning the triplet query into a 500 response.

This is reachable through the event POST endpoint, whose event data is intentionally open-ended, and through OpenAI-compatible responses that do not echo prompt token IDs. _trim_model_request also defaults missing IDs to an empty list.

Minimal reproduction

1. Append two model_request events to one rollout with different non-empty response token IDs.
2. Give both events prompt_token_ids: [] (or omit the field).
3. Query /api/rollouts/{id}/events?format=triplet.

Only the second request is returned. Using prompt_token_ids: [[1]] instead raises TypeError: unhashable type: 'list'.

Expected behavior

Prompt-based last-wins deduplication should remain unchanged when IDs are a valid, non-empty list of integers. When no valid key exists, the events should be retained in insertion order because the value cannot establish that they are duplicates.

I have a focused fix with server and VERL multimodal-alignment regression coverage ready and will link the PR here.