Normalize observation-only steps without blank agent messages

[neubig]

Some standardized samples represent environment/tool observations as blank source: agent steps with only an observation field. This is schema-valid, but semantically awkward because the environment, not the agent, authored the observation.

Examples from the audit/heuristics include web/mobile datasets such as WebLINX, Wonderbread, Mind2Web, Go-Browse-WA, TurkingBench, WebArena, AndroidControl, and related browser/mobile traces.

Acceptance criteria:

• Decide the canonical representation for standalone observations when no preceding tool call is available.
• Prefer attaching observations to the triggering tool-call step when possible.
• If standalone observation steps remain necessary, document the convention clearly or add schema support for environment-authored steps.
• Regenerate affected samples after the representation is settled.