Fix: recover lossy TOOL-span args from the LLM tool_call (LangChain/LangGraph)

[AshwinUgale]

What

The OTel/OpenInference adapter reads a tool call's arguments from the TOOL
span's input.value. LangChain / LangGraph records that as a lossy bare
scalar — e.g. "A100" instead of {"order_id": "A100"}. The args then parse as
empty, and on a perfectly valid call:

• R6 (malformed_arguments) fires a false hard_defect, and
• R1 (schema) sees no args and reports the required field missing,

both hard_defects → the run fails CI (exit 2).

The model's real, valid arguments are in the trace the whole time — on the
originating LLM span's tool_calls.*.tool_call.function.arguments
({"order_id": "A100"}). The adapter just wasn't using them when a TOOL span
was present.

Fix

Index each LLM span's output tool_calls by tool name (in span order). When a
TOOL span's own input.value does not parse to an argument object, recover the
arguments from the matching LLM tool_call and clear raw_text (the arguments
were never malformed — the TOOL span just under-recorded them).

• Instrumentors that already put full args on the TOOL span (smolagents:
  {"args":[],"kwargs":{...}}) are unaffected — recovery only runs when the
  TOOL span's own args are empty.
• A genuinely malformed model call yields no structured args on the LLM span
  either, so R6 still fires in that case (only output_messages — the newly
  emitted calls — are read, so a replayed turn never double-counts).

Why it matters

Second per-instrumentor shape difference found while running tracelint against
real third-party OpenInference traces (after the nested content-parts fix,
#52). Instrumentors place the authoritative tool arguments in different spans
— smolagents on the TOOL span, LangChain on the LLM span — and the shared OTel
adapter must normalize that. Without this, any LangGraph/LangChain user gets a
spurious CI failure.

Verification

Built from a real gpt-4o-mini LangGraph create_react_agent trace. Before:
R6 (keyless) and R1 + R6 (with tools.json), both exit 2. After: 0
findings, exit 0 in both cases; R1 coverage 1/1. smolagents real traces
unchanged.

Tests

tests/test_adapter_otel.py::test_lossy_tool_span_input_recovers_args_from_llm_tool_call
(built from the real LangGraph shape). ruff check + full suite green.