Adapter: record CHAP human-decision events from LlamaIndex Workflows human-in-the-loop

[BrightbeamBogdanCastraveti]

Closes #1.

Adapter mirroring chap-langgraph / chap-pydantic-ai: when a LlamaIndex Workflow step pauses on an InputRequiredEvent, the human's HumanResponseEvent is recorded as a CHAP decision.

approve   -> decide.approve
override  -> decide.override   (proposed vs returned, as an RFC 6902 diff)
reject    -> decide.reject

Design (per the plan on #1):

• decision is explicit and authoritative — Workflows events are schemaless and carry no approve/edit/reject signal, so the adapter never infers intent. proposed/returned/rationale/tags/intent_preserved/approver default to event.get(...) with per-call overrides; arg names match the sibling bridges.
• Recorded from the driver (where both events meet), using the two-step InputRequiredEvent → HumanResponseEvent loop — so step replay never double-records.
• llama-index-workflows is optional; the bridge reads events structurally, so it and its tests run without it.
• 0.2.6-correct by construction: approver joined, review.request addressed to that approver, so from is always a member and in the to set.

Verified against main (0.2.6): adapter pytest 12 passed; example runs on real llama-index-workflows 2.22 through approve / refining edit / substituting edit / reject; conformance harness 23/23.

[BrightbeamBogdanCastraveti]

Self-review caught one edge case worth flagging: in record_decision, if decision="override" but no returned content is supplied (neither the returned= argument nor a response field on the event), _diff(proposed, None) produces a whole-artefact replace with null — i.e. it records a blank-everything override instead of refusing. Tests and the example always pass returned, so it doesn't surface there, but a malformed caller could write a misleading record.

Fix is a one-line guard — raise instead of diffing against None:

if diff is None:
    if returned is None:
        raise ValueError("an override needs `returned` content (or a `diff`)")
    diff = _diff(proposed, returned, "")

Ready locally; I'll push it once the branch-update ruleset is scoped (same blocker as #5).

[brightbeamarsalan]

you should be able to update Bogdan, please let me know how it goes.

[BrightbeamBogdanCastraveti]

Pushed c2498ee — the branch updated cleanly now that the ruleset is scoped. That's the override guard (raise instead of a null-replace when an override has no returned content), with a test. Re-ran against main (0.2.6): pytest 14 passed, harness 23/23. Ready for review.