[codex] Map AI failure spans to error taxonomy

[rickcrawford]

Summary

• Map AI request span failures to the low-cardinality error taxonomy from WOR-1215: guardrail_blocked, rate_limited, content_filter, budget_exceeded, upstream_5xx, and timeout.
• Record ERROR status and error.type on the actual ai.request span for rate limits, guardrails, budget preflight failures, provider status failures, streaming read failures, and all-provider transport failures.
• Align observability docs and regenerated docs/llms-full.txt with the taxonomy.

Root cause

Some AI failure paths recorded on tracing::Span::current() even though the ai.request span was not entered, and provider response failures were not consistently classified for span status/error attributes.

Validation

• cargo test -p sbproxy-core ai_error_classification_tests --locked
• cargo test -p sbproxy-ai record_error_marks_span_failed --locked
• cargo test -p sbproxy-ai error_type_constants_cover_ai_failure_taxonomy --locked
• cargo test -p sbproxy-core --locked
• cargo test -p sbproxy-ai --locked
• cargo fmt --all --check
• git diff --check
• ./scripts/regen-llms-full.sh --check
• cargo test -p sbproxy-e2e --test otlp_span_arrival_e2e --locked
• cargo test -p sbproxy-e2e --test otlp_span_arrival_http_e2e --locked

No new libraries were added; NOTICE did not need changes.

Linear: WOR-1215