test: backend coverage parity audit post-refactor #1054

Background

Refresh of retired sub-issue #477. After the plugin refactor, verify all backends emit spans uniformly and parity holds across openai, ollama, huggingface, watsonx, litellm (and vllm if applicable).

Scope

• Add integration tests (gated on backend-specific markers, per test/MARKERS_GUIDE.md) asserting each backend produces backend-generation spans with a matching set of gen_ai.* attributes.
• Document any backend-specific attribute gaps.
• Cross-check the Anthropic-specific cache/reasoning token attributes emitted by PR feat(telemetry): close five OTel GenAI semantic convention emission gaps (#1035) #1036 (gen_ai.usage.cache_read.input_tokens, gen_ai.usage.cache_creation.input_tokens, gen_ai.usage.reasoning.output_tokens) against the current OTel GenAI semconv at audit time; rename to match the spec if equivalents have been standardised, otherwise flag with a code comment as Anthropic-specific extensions pending standardisation. (Per @planetf1's comment on Epic: Enhanced distributed tracing #444.)

Phase & dependencies

Phase 3 (polish). Depends on #1045 and #1046. Can start any time after Phase 1 lands; best done after most of Phase 2 has merged so the audit is against steady-state tracing.

Acceptance criteria

• All backends produce backend-generation spans with a shared minimum attribute set.
• Known gaps documented.
• Anthropic-specific cache/reasoning token attribute names cross-checked against OTel GenAI semconv; renamed or flagged as documented above.

Parent epic: #444