[OTel] _flatten_nested_dict bypasses the shared decoder, so nested-format traces can still produce unhashable tags

[Leroyyyyyyyyy]

Description

Split out of the #187 review, as agreed there.

OtlpJsonLoader._extract_attributes has two branches, and only one goes through the shared AnyValue decoder:

def _extract_attributes(self, attrs) -> dict:
    if isinstance(attrs, dict):
        return self._flatten_nested_dict(attrs)   # no decoding, no allowlist
    return decode_attributes(attrs)               # decoder + SPEC_CONTAINER_ATTRS

The dict branch handles the flat and nested (ClickHouse JSON column) formats.
It recurses into dicts and passes everything else through untouched:

for key, value in d.items():
    full_key = f"{prefix}{key}" if not prefix else f"{prefix}.{key}"
    if isinstance(value, dict):
        result.update(OtlpJsonLoader._flatten_nested_dict(value, full_key))
    else:
        result[full_key] = value       # a list lands in span.tags as a list

So the guarantee #187 establishes on the OTLP-array path — containers only for allowlisted keys — does not hold here. A list reaches span.tags, and gcp.vertex.agent.invocation_id is then used as a dict key in eight places in incremental_processor.py (llm_spans_by_invocation, token_totals, seen_tool_calls).

This predates #187 — the dict branch never had any narrowing — so it is not a regression from that PR, but it is the same hazard on a path the PR does not cover.

Expected behavior

Both branches of _extract_attributes should end up with the same contract: container values only for keys on the allowlist, narrowed otherwise. A nested-format trace should not be able to put an unhashable value where the OTLP-array format cannot.

Steps to reproduce

Load a nested-format (ClickHouse JSON column) trace whose attributes carry a list where a scalar is expected:

{"gcp": {"vertex": {"agent": {"invocation_id": ["a", "b"]}}}}

_flatten_nested_dict flattens it to {"gcp.vertex.agent.invocation_id": ["a", "b"]}, which reaches self.seen_tool_calls[invocation_id] in incremental_processor.py and raises TypeError: unhashable type: 'list'.

How are you using agentevals?

CLI file evaluation (agentevals run)

Config dump (ZIP from the web UI)

No response

Version information

agentevals version: 0.9.9.dev10+gc65f82d19.d20260828
Python version: 3.14.6
OS: macOS 26.4 (arm64)

Eval config (if applicable)

Trace and eval set files

No response

Relevant logs or error output

Additional context

Likely fix: apply the same allowlist rule in _flatten_nested_dict's leaf case, via a shared helper rather than a second copy of the rule.

I can take this one if it's useful — same area as #187.

Human confirmation

• I am a human (not a bot, agent, or AI) filing this issue.