fix(openai): align first-chunk fields and cover response combinations

[arekay-nv]

Mixed OpenAI chat deltas can contain content, reasoning, and tool calls in the same first response chunk. TPOT used to infer which field held the first token from the final fields, which is only right when a response starts with pure reasoning. #526 kept reasoning from mixed deltas (fixing OSL) but kept that inference, so TPOT counted content already delivered with the first token, or dropped tokens that arrived later:

Stream (delta order)	Correct TPOT text	Before (#526)	This PR
reasoning → content	R1 C0 C1	✅	✅
reasoning + content in first delta	C1	❌ C0 C1	✅
content → reasoning	R0 C1	❌ C0 C1	✅
content + tool start → tool tail	tool tail fragment	❌ whole tool call	✅
tool start → content	C0 C1	❌ C1 + first tool fragment	✅

Align entry zero of each streamed TextModelOutput field with the first meaningful delta. Fields that start later get an empty first entry; text tails are aggregated and tool-call batches retain their indices for merging. The accumulator keeps a reference to the first delta alongside its existing contributions. Full-output consumers combine all entries, while TPOT skips entry zero independently in every field. No new output class or wire field is added.

Existing event logs keep their original TPOT on replay. Earlier revisions stored reasoning responses as a joined output string beside streamed reasoning. The current accumulator never writes that shape, so TextModelOutput and the steady-state diagnostics count all of that string as tail, matching the TPOT those runs reported. This covers scripts/early_stopping_estimate_from_events.py and the diagnostics CLI. Neither replays tool-call samples, so the older tool-call layout needs no separate rule.

Adds 292 response-contract test cases, including:

• All eight content/reasoning/tool combinations across two deltas, both reasoning aliases, and both streaming modes (256 cases).
• Single-delta output, empty/role-only/null/control messages, interleaved tool arguments, Unicode, and JSON/MessagePack round trips.
• Hypothesis-generated sequences and chunk splitting after a fixed first delta.
• Real-tokenizer OSL/TPOT checks through the production event codec for mixed-first, tools-before-text, content-before-reasoning, and single-mixed-delta responses.

Also covers replay of the joined-output layout through TextModelOutput and the offline diagnostics, and updates existing fixtures and offline diagnostics to the aligned-field contract. Follow-up to merged PR #526.

Validation:

• 852 tests passed across OpenAI adapters, core serialization, metrics aggregation, metrics and steady-state diagnostics, and the early-stopping replay script. One existing Linux-only CPU-affinity test was deselected on macOS; a short temporary directory avoids the macOS IPC path limit.
• Replaying a pre-change reasoning event log through scripts/early_stopping_estimate_from_events.py gives the same TPOT as the revision that produced it (1111 ns; 2500 ns without the replay rule).
• All pre-commit hooks passed. mypy targeting Linux reports no issues (335 files); the default macOS target reports the three existing errors for Linux CPU-affinity APIs.
• Synthetic response inputs and the checked-in hermetic chat tokenizer; no live GPU endpoint runs.

🤖 Generated with Claude Code

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅

[leopck]

LGTM