feat(instrument): OpenTelemetry GenAI semantic conventions for all LLM-call adapters (spec 07)

[mmercuri]

Summary

Implements spec 07-otel-genai-semantic-conventions.md — every LLM-call event emitted by every adapter (9 providers + 16 framework + embedding) is additively stamped with the canonical OpenTelemetry GenAI gen_ai.* attribute set, alongside the existing custom LayerLens attributes.

• New shared helper: src/layerlens/instrument/adapters/_base/genai_semconv.py (constants + stamp_genai_attributes + detect_gen_ai_system)
• Centralised hook in BaseAdapter.emit_dict_event covers every framework adapter automatically
• Per-adapter explicit stamping in LLMProviderAdapter._emit_model_invoke and _emit_tool_calls for accurate gen_ai.system / gen_ai.tool.* resolution
• Stamping is idempotent, additive (no legacy field removed), and safe (never raises on partial payloads)
• Documentation: docs/adapters/otel-genai-conventions.md — contract + per-adapter coverage matrix

gen_ai.* Attribute Set Stamped

Required (always present on model.invoke):
gen_ai.system, gen_ai.provider.name, gen_ai.operation.name

Request: gen_ai.request.{model, max_tokens, temperature, top_p, top_k, frequency_penalty, presence_penalty, stop_sequences, seed, choice.count, encoding_formats}

Response: gen_ai.response.{id, model, finish_reasons}

Usage: gen_ai.usage.{input_tokens, output_tokens}

Tool: gen_ai.tool.{name, call.id, description, type}

Agent: gen_ai.agent.{id, name, description}

Provider-specific:

• OpenAI: gen_ai.openai.{request.service_tier, request.response_format, response.service_tier, response.system_fingerprint}
• Anthropic: gen_ai.anthropic.{cache_creation_input_tokens, cache_read_input_tokens}
• AWS Bedrock (per spec §4.3 namespace): aws.bedrock.{guardrail.id, knowledge_base.id, agent.id}
• Google Vertex: gen_ai.google.safety_ratings

Per-Adapter Wiring Coverage

Layer	Adapters	Wiring path
Providers (9)	OpenAI, AzureOpenAI, Anthropic, AWSBedrock, GoogleVertex, Cohere, Mistral, Ollama, LiteLLM	explicit stamp in _emit_model_invoke + centralised hook
Frameworks (16)	agno, autogen, bedrock_agents, crewai, google_adk, langchain, langfuse, langgraph, llama_index, ms_agent_framework, openai_agents, pydantic_ai, semantic_kernel, smolagents, strands, salesforce_agentforce	centralised hook (every emission goes through BaseAdapter.emit_dict_event)
Embedding	EmbeddingAdapter (embedding.create → gen_ai.operation.name=embeddings)	centralised hook

Test plan

• uv run pytest tests/instrument/adapters/_base/test_genai_semconv.py -x — 49 passed
• uv run pytest tests/instrument/adapters/test_genai_semconv_per_adapter.py — 35 passed
• uv run pytest tests/instrument/adapters/frameworks/ — 258 passed, 1 skipped (no regression)
• uv run mypy --strict src/layerlens/instrument/adapters/_base/genai_semconv.py — clean
• uv run ruff check — clean

CLAUDE.md Compliance

• All LLM-call adapters wired (no "planned backlog" deferrals)
• gen_ai.* attribute names match official OTel spec verbatim (constant table is the source of truth; TestGenAiAttributeNamesMatchSpec pins each one)
• Additive contract preserved — no legacy attribute removed; tests assert side-by-side coexistence
• No co-author trailers