fix(codex): preserve commentary across interleaved output items

[ruizanthony]

Summary

• preserve commentary buffers when Codex Responses output items interleave
• keep unmatched aliased deltas private while a private item is active
• preserve standalone unphased final output streams instead of dropping visible text

Context

PR #59831 was superseded by #66115, which landed the core commentary-to-interim routing and display.show_commentary control. This is a narrow follow-up for residual stream-shape cases reproduced against current main; it does not reintroduce the superseded implementation.

Tests

• HERMES_PYTHON=/usr/local/lib/hermes-agent/venv/bin/python ./scripts/run_tests.sh tests/run_agent/test_run_agent_codex_responses.py -q — 112 passed
• python -m py_compile agent/codex_runtime.py
• git diff origin/main..HEAD --check

[ruizanthony]

Maintainer action needed: the CI workflows on this PR are awaiting approval (first-time contributor gating) — no checks have run yet on head 780aefe72a. Could you approve the workflow run so the gate can certify this head? Local gate on the contributor side: codex responses/streaming regression suites pass (42/42 on the rebased patch). Thanks!

[ruizanthony]

Rebased onto current master (4d9541b9c) — the PR is mergeable again. Semantic content unchanged: the 3 commits are preserved (interleaved commentary items, unmatched aliased deltas privacy, unphased final streams); the only delta vs the previous head is upstream context adaptation (event_stream → concrete_response rename in the concrete-response compatibility block). Combined diff verified line-by-line against the pre-rebase patch: no behavior loss. tests/run_agent/test_run_agent_codex_responses.py: 41/41 passed. New head: 100799377. Still needed from a maintainer: approval of the pending workflow runs (first-time contributor gating) so CI can certify this head.

[teknium1]

Thanks for the focused follow-up. The premise remains valid on current main: agent/codex_runtime.py:1069-1078 stores only one active message phase, and agent/codex_runtime.py:1083-1101 routes deltas solely through that mutable value. That cannot preserve a commentary item when another output item interleaves before its remaining deltas arrive.

The proposed item-id/output-index association addresses that state-model gap, and the added regression coverage exercises interleaved commentary/final items, function-call interleaving, alias mismatches, and the Relay concrete-response path. I found no blocking correctness or design-fit issue in the diff.

Automated hermes-sweeper review.