perf: avoid returning unused thread history during session resume

[robinfai]

Summary

• pass excludeTurns: true when ACP session/resume and session/load establish the app-server subscription
• keep thread/read({ includeTurns: true }) unchanged for session/load, preserving full ACP history replay
• add a narrow local type for the experimental resume field and regression coverage for both lifecycle paths

Why

thread/resume includes the reconstructed turn history by default. Neither ACP path consumes that response history: session/resume must not replay it, while session/load deliberately reads the authoritative full history with a subsequent thread/read.

In one observed run on a real legacy thread backed by a 486 MiB rollout (77 turns, 2,306 items), Codex 0.147.0 produced:

Resume request	Response size	Response time	Client JSON parse
default	65,991,086 bytes	1,332.77 ms	32.61 ms
excludeTurns: true	1,461 bytes	1,079.24 ms	0.02 ms

This removes one unused large JSON response from both lifecycle paths. For session/load, the required full history still comes from thread/read, so the history and tool-call replay behavior introduced in #208 remains unchanged.

The adapter already opts into app-server experimental APIs and depends on @openai/codex ^0.147.0. The checked-in generated bindings currently omit experimental fields, so this uses the same local-type pattern already used for experimental thread settings instead of regenerating the whole experimental schema.

Validation

• npm run typecheck
• npm test — 357 passed, 28 skipped, including a transport-level assertion for the experimental JSON-RPC parameter
• npm run build
• npm run codex-test -- -p "Reply with exactly: resume-history-validation" -o summary — completed with end_turn (49 Codex events, 8 ACP events)
• direct Codex 0.147.0 app-server comparison on an isolated copy of the real rollout