fix(core): normalize AI SDK fragment boundaries

[rekram1-node]

Before

AI SDK providers may overlap, repeat, or omit text and reasoning boundaries. The AI SDK adapter forwarded them unchanged, but Session accepts one open text fragment and one open reasoning fragment at a time, so these streams killed the turn with an untyped defect.

In #50662, AI SDK's OpenAI Responses parser behind Vercel AI Gateway starts summary part :1 while :0 is still open, then ends both at item completion:

reasoning-start(rs_1:0)
reasoning-start(rs_1:1)   -> "reasoning start before end"

Native @opencode/ai protocols already normalize these boundaries through their lifecycle helpers (#45789). Providers that run through the AI SDK adapter did not.

After

The adapter applies the same rule to text and reasoning: a start or delta for another fragment closes the open one, repeated starts are ignored, and ends for fragments that are not open are dropped.

reasoning-start(rs_1:0)
reasoning-end(rs_1:0)
reasoning-start(rs_1:1)

Well-formed streams are unchanged. This also covers the same class of failure in other AI SDK producers:

• repeated reasoning starts from older xAI Responses
• reasoning deltas after an ended id from xAI Chat (reopened, not dropped)
• message items that stay open until flush in xAI Responses
• the same summary-part overlap in the in-repo Copilot Responses fork

Dropping a late end for a force-closed part loses only that end's metadata. OpenAI repeats the same encrypted content on every summary-part end, so the part that remains open still carries it for replay.

Testing

• bun test test/aisdk.test.ts, test/github-copilot, and test/session from packages/core
• bun typecheck from packages/core
• bun run check

Closes #50662