fix: normalize BYOK responses usage

[tharuxpert]

Summary

• Normalize Chat Completions usage into Responses-style input_tokens, output_tokens, and total_tokens for BYOK routes.
• Preserve cache/accounting details under input_tokens_details / output_tokens_details when upstreams provide them.
• Capture Anthropic message_start.usage so stream cache usage is not dropped.
• Ignore local codedb.snapshot inspection artifacts.

Why this matters

Some OpenAI-compatible BYOK providers return Chat-style usage (prompt_tokens / completion_tokens). Codex expects Responses-style usage on response.completed; when the shape is wrong, a successful model response can still make Codex reject the turn and retry the request. This fix keeps BYOK agent loops usable and preserves token/cache accounting for compaction and cost visibility.

Verification

• python3 -m pytest tests/ -q
• python3 -m compileall codex_shim/ -q
• codex-cli 0.134.0 through codex-shim to mimo-v2.5-pro: turn completed successfully with cache usage reported (input_tokens: 25728, cached_input_tokens: 25664).

Security review

Diff-scoped codex-security review completed. No exploitable issue found in the changed code path.