refactor(ai): resolve responses item ids once at the stream boundary

[rekram1-node]

Why

The Responses parser had grown a second identity path for function_call items that arrive without an id: item.id ?? call_id coalescing, Object.values(state.tools).some((tool) => tool?.id === call_id) scans on dedup and terminal reconciliation, conditional itemId metadata, and per-event "message is missing id" guards. #46084 extended it further. The parser body should be able to assume item.id exists; the missing-id policy belongs in one place.

What the wire actually does

Probed every Responses API we have credentials for, with a plain-text turn and a forced tool-call turn each:

Provider	output_item.added	output_item.done	response.completed.output
OpenAI (gpt-5-mini)	id	id	id
Azure (gpt-5.4-nano)	id	id	id
xAI (grok-4-1-fast*)	id	id	id
Meta (api.meta.ai, muse-spark-1.1)	id	id	id
Perplexity (perplexity/sonar, openai/gpt-5-mini)	id	id	id
Bedrock Mantle (openai.gpt-oss-120b/20b)	id	no id — renamed to item_id, stray output_index inside the item	no id — same

So the Open Responses schema is right that id is required, and five of six providers honor it, but Mantle's gateway drops it from terminal function_call items. That is the real case the old fallbacks were serving, and a strict schema would regress it.

What changes

• One resolution point. OpenResponses.normalize(state, event) runs before step and does two things: the existing output_index → item_id remap for delta events, and item-id resolution mirroring Codex's assign_missing_streamed_response_item_id: an item without an id adopts the id already registered in its output slot (state.outputItems[output_index]; for response.completed.output, the array position), otherwise it gets a locally minted fc_/msg_/rs_/cmp_ + UUID. step, onOutputItemAdded, onOutputItemDone, and the hosted-tool helpers take OutputItem = StreamItem & { id: string } and never look at a missing id again.
• Removed the call_id scans, item.id ?? call_id coalescing, conditional itemId metadata, isReasoningItem narrowing, and the message is missing id guards. Pending tools, completedTools, and terminal reconciliation are keyed by item.id only.
• Provider steps (openai-responses, xai-responses) call normalize first; outputItemID is no longer exported.
• Tests: added a fixture captured from the real Mantle gpt-oss stream (id missing on done/completed, item_id present, stray output_index) and a done-only no-id case asserting a minted fc_ id. Removed tests whose only purpose was id-less items; fixed fixtures that omitted ids incidentally.

Live verification

Ran a real two-turn tool-call conversation through the patched client, switching provider between the tool call and the tool result:

First turn → second turn	Item id on replay	Result
OpenAI → OpenAI, Azure → Azure, xAI → xAI	real server-issued id	ok
Mantle → Mantle	fc_… recovered from output_item.added (the Mantle case)	ok
OpenAI → Mantle, Azure → Mantle	real OpenAI-style fc_… replayed to Mantle	ok
Mantle → OpenAI, Mantle → Azure, xAI → OpenAI, xAI → Azure, OpenAI → Azure	dropped (provider metadata key differs)	ok
minted fc_<uuid> → OpenAI / Azure / xAI / Mantle	minted id on the wire (confirmed via compileRequest)	all accepted

cd packages/ai
bun typecheck
bun run test   # 979 pass, 28 skip, 0 fail