Emit hydrated spans for better trace inspection UX

[adagradschool]

Summary

This updates the Claude trace plugin so Claude LLM spans contain the full prompt history for the model call instead of only the incremental delta since the previous call.

It also simplifies the hook lifecycle by making Stop synchronous and removing the extra Stop/SessionEnd coordination state.

What changed

• parse_stop_transcript.py

  • rebuilds prompt history from the full transcript
  • still emits spans only for the latest turn
  • adds promptlayer.prompt_history_mode = "full_session"
  • dedupes duplicate user entries caused by queue/transcript overlap

• Hook flow

  • makes Stop synchronous
  • removes async-era session state used to coordinate Stop and SessionEnd
  • keeps SessionEnd focused on final root-span emission and cleanup

• Tests

  • adds a multi-turn full-history fixture
  • adds parser and OTLP assertions for full-session prompt emission
  • extends fixture replay coverage

Why

This makes Claude span data reflect the full model-call context, which improves trace inspection for multi-turn and tool-using sessions.

Verification

• bash scripts/replay-fixtures.sh
• source .env && uv run --script scripts/e2e_smoke.py

[rite2hhh]

LGTM