[harryrberg]

Codex models can advertise Ultrafast, but ACP clients currently receive only a Fast on/off control. Add an Ultrafast value to the existing fast-mode option and use Standard/Fast/Ultrafast display names for select clients.

• Discover support from serviceTiers or legacy additionalSpeedTiers, without hardcoded model names.
• Keep the existing ID, on/off values, and boolean toggle for models without Ultrafast. Models with Ultrafast use a select, including for boolean-capable clients.
• Send the selected tier on prompts, recognize it on new/resumed/loaded/forked sessions, and fall back to Fast or Standard when a new model cannot use Ultrafast.
• Cover ACP v1/v2, capability filtering, invalid choices, model changes, and consecutive prompts. No personal model, reasoning, or speed defaults are included.

Related: #673 preserves arbitrary configured tiers, including Flex. This PR focuses on explicit client selection and overlaps with its session-state/type handling; the two changes will need reconciliation. #648 also overlaps with catalog Fast capability detection. Generated protocol files are unchanged.

Validation on Windows / Node 24:

• Typecheck and build passed.
• 37 focused tests passed across four test files.
• Built adapter: selected GPT-6.1 Sol, Low, Ultrafast through ACP and completed a real prompt (OK, end_turn). The repository run-codex helper also completed a separate default-model smoke prompt. These are functional checks, not throughput guarantees.
• A broader suite run at the preceding revision reported 2,284 passed, 70 failed, 47 skipped. Representative failures in CodexAcpClient, diff-v2, and session-index-v2 also reproduce in an untouched upstream worktree (13 failures across those three files). The complete suite is not claimed green. Unrelated Windows snapshot rewrites were discarded.

Opening as a draft while upstream overlap and cross-platform CI are reviewed.