Based on this mailing-list question and discussion in #4381, I think we have an option to include the choice sequence and spans in observability outputs (respectively as metadata.choices and metadata.choice_spans).

• The choice sequence can mostly be represented directly in JSON, plus {"nan": <integer value>} and {"bytes": <base64 value>}. And {"integer", <string value>} for large integers...
• We might want to store the choice nodes though, which have the constraints on each choice? Hmm. metadata.choice_nodes if so.
• Spans also seem useful, though it's not entirely clear how to store them - maybe a dict with keys label/start/end, and lists of integers?
  • We should probably also ship functions to transform these into various forms though, including (choice_sequence, spans) -> labelled_tree.

Finally, I'm inclined to say that this means observability is graduating from experimental to a normal supported part of Hypothesis; and we should therefore think a bit about how it's configured. We'll have options (none, basic, coverage, choices, all); might we end up adding more? If not, it's easy to make that a setting and HYPOTHESIS_OBSERVABILITY= env var.

• (liam:) get tyche to recommend the new stable observability interface in its docs :)