fix(instrumentation-py): repair SageMaker, Semantic Kernel, and Smolagents OTel 2.x spans

[Nightingalelyy]

Summary

• repair SageMaker streaming, async-job, provider, canonical tool, bounded serialization, and shared lifecycle behavior
• normalize Semantic Kernel prompt wrappers, current chat operations, canonical tool definitions/calls, stable tool I/O, privacy, and shared lifecycle behavior
• normalize Smolagents tool/step/agent spans, canonical tool data, bounded privacy-safe serialization, and shared delegate lifecycle behavior
• add the patch release intent for exactly these three Python instrumentation packages

Validation

• Read every current document under respan/contribution/ and audited the complete source and paired example diffs against the authoring, span-contract, architecture, testing, CI, publishing, and release-intent rules
• Focused package tests pass: SageMaker 9/9, Semantic Kernel 17/17, Smolagents 19/19 (45/45)
• Root-scoped Ruff check and format-check pass for all changed Python package files
• All three package wheels build; clean-target import smoke passes, including final Semantic Kernel wheel SHA-256 1dbd262a18d293947a6f561ea86e8797a0c01505a2bd2bdf703e588d5633f3d0
• Release inventory, release-intent validation, missing-intent check, and git diff --check pass
• All 12 paired examples pass against exact local editable packages under marker otel2-fix-py-group-25-20260818T071900Z
• Respan MCP audit opened all 12 trace trees and all 44 full records; every tree has one root, exact parentage, matching counts, and unique log/span IDs
• Full-record audit confirms exact marker/grouping on 44/44 records, complete bounded inputs/outputs, exact model usage, SageMaker streaming, canonical tool definitions/current-turn calls, stable tool I/O, no forbidden aliases, and no credential or object-address leakage

External follow-ups

• Platform ingestion still renders native failed Semantic Kernel and Smolagents spans as success/200 despite retained error fields and canonical HTTP 500 metadata.
• Platform convenience projections still report SageMaker model traffic as provider openai despite canonical gen_ai.system and gen_ai.provider.name being sagemaker, and top-level tools/tool_calls arrays remain empty despite canonical attributes.
• Semantic Kernel task display and Smolagents agent model labels remain backend/UI derivations; the instrumentation follows the contribution contract and does not add forbidden aliases or LLM fields to common-only Agent/Task spans.
• A real SageMaker endpoint remains credential/service-gated because no SAGEMAKER_ENDPOINT_NAME is configured; committed deterministic Stubber paths cover text, tools, response streaming, and async submission.

Companion examples

• Paired example PR: fix(examples-py): validate SageMaker, Semantic Kernel, and Smolagents OTel 2.x spans respan-example-projects#73

[greptile-apps]

Your trial has ended. Reactivate Greptile to resume code reviews.