[codex] Fix hook flush overhead

[jverre]

Summary

• Skip synchronous context snapshot construction when a flush batch contains only tool spans, avoiding full-transcript and filesystem scans that cannot be attached to any LLM usage span.
• Reuse the already-read turn transcript when computing close-out edit metrics, removing one duplicate transcript pass on Stop/SessionEnd.
• Treat non-2xx count_tokens responses as errors so the detached measurement pass does not cache zero-token anchors from auth/rate-limit/error payloads.

Review notes

Recent changes added useful billing and attribution detail, but the hook hot path had one avoidable cost: flush always called buildContextSnapshot, even for tool-only batches. That helper reads the full transcript and hits memory/agent/skill/tool extractors; doing that during ordinary tool hook flushes slows Claude Code without adding metadata because only spans with usage receive the snapshot.

I did not include the local bin/opik-logger-darwin-arm64 modification in this PR; it was already dirty before the source edits and looks like a local build artifact.

Validation

• go test ./...
• go vet ./...
• git diff --check