[etserend]

Problem

When an app re-sends the whole conversation (as graph state or as a message list), every agent and workflow span
exported all earlier turns in gen_ai.input.messages. Each new trace's input repeated the previous inputs and outputs,
and O11y Cloud rendered them as one block that grows with the conversation. The output side already trims history; the
input side did not.

Change

_set_orchestration_content now trims the exported input to the messages from the last user message onward, when the
input is either a {"messages": [...]} state container or a bare list of role-tagged messages (the shape the LangChain
agent middleware logs). Tool-call and tool-response messages of the current turn are kept. The untrimmed input is still
used for the existing output history dedup.

Unchanged: partial update containers, single-turn inputs, inputs with no user message, and LLM (chat) span inputs
(see Notes).

Verification

Multi-turn runs (3 turns, the full history re-sent each turn), spans captured on a local collector. Message counts of
the root agent or workflow span input per trace, before → after:

Framework	Root input	Result
LangGraph via the callback handler (healthcare assistant example)	1 / 3 / 5 → 1 / 1 / 1	fixed
LangChain create_agent + SplunkAOMiddleware	1 / 5 / 9 → 1 / 1 / 1	fixed
OpenAI Agents SDK	1 / 1 / 1 → 1 / 1 / 1	unaffected (string inputs)
CrewAI	not run	the dependency mirror had no chromadb wheel for macOS arm64

Inner chatbot, tools_condition and tools spans of the LangGraph run now show only their own turn, and the lab0
session view shows each question once. Chat span inputs are identical before and after in every run.

• tests/test_attribute_mapping.py: 75 passed, including 15 new cases (behaviours and guards, with most run for both
  workflow and agent spans). The cases for the first change failed before it.
• ruff check, ruff format --check on the changed files, and mypy (invoke type-check) pass.
• The full suite shows 24 failures and 5 errors in exporter diagnostics, middleware tracing, OpenAI/OpenAI Agents,
  serialization and two collection errors. The identical set fails on unmodified origin/main, so none come from
  this change.

Notes

• Behaviour change: earlier turns are no longer in workflow and agent span inputs; they remain in the session and in
  LLM span inputs.
• Chat spans keep the full prompt. An LLM span's gen_ai.input.messages records the exact messages the framework
  sent to the model (system prompt, earlier turns, tool calls and results), so it is not trimmed. In the
  verification run, the third turn's chat span input held 6 messages (system, then the whole conversation), and its
  follow-up call after the tool call held 8. Only agent and workflow spans carry app state rather than a prompt, so
  only they are trimmed. This is how the LLM span path already worked (set_llm_attributes writes span.input as
  is), and this PR does not change it.
• Trade-off: a bare message list cannot be told apart from conversation state, so a workflow or agent input that is
  deliberately a message list (for example few-shot examples or a system prompt) is trimmed to the current turn too.
• The LangChain middleware path has no working unit test in the repository: tests/test_langchain_middleware.py
  already errors at collection on origin/main. The live run above is the evidence for that path.
• CHANGELOG.md updated under Unreleased.