fix(responses): preserve valid replay item graphs (#97427)

[lgy1027]

Fixes #97427.

When replaying encrypted reasoning with store: false, the server-side reasoning ID is intentionally omitted. This change also omits the paired assistant-message ID, preventing the API from rejecting an incomplete replay graph.

It preserves short message IDs for turns without replayed encrypted reasoning and for messages after a reasoning-to-tool round, retaining cache reuse where the graph is independent.

Tests cover:

• paired reasoning/message replay
• plain assistant-message replay
• direct preflight normalization
• a reasoning → tool-call → tool-output → independent-message boundary

Validation:

• uv run --extra dev pytest tests/agent/test_codex_responses_adapter.py -q
• uv run --extra dev ruff check agent/codex_responses_adapter.py tests/agent/test_codex_responses_adapter.py

[alt-glitch]

This was generated by AI during triage.

Duplicate of #97445. Both repair the same Responses replay invariant: an assistant message ID must be removed when its encrypted-reasoning ID is stripped. #97445 is the earlier, broader open fix.