Expose per-session cost to hooks / TUI (or add the a cost session section in the TUI)

先追蹤 PreToolUse、PostToolUse、Stop 和 SessionStart 的 hook payload，接著檢查工作階段 transcript JSONL，以及 issue 中提到的現有 /internal/usage/summary 路徑。確認每一輪的使用量和模型資料是否能透過 hooks 暴露、持久化至 transcripts，或針對某個工作階段回傳；完成的標準是存在一種受支援的方式，可取得目前工作階段的成本以供 TUI 顯示。

Feature Description

Today hooks (PreToolUse/PostToolUse/Stop/SessionStart) receive a stdin payload with tool_input/tool_response, but it does not include the
LLM usage (input/output/cache tokens) or the model per turn. Also, the session transcript JSONL
(~/.commandcode/projects//.jsonl) persists only the message content, not the usage.

I know CommandCode already computes usage internally (it powers the "updated" indicator in the TUI) and exposes /internal/usage/summary at
the account level, but there's no way to see the accumulated cost of the current session from outside.

Request: include usage (input_tokens, output_tokens, cache_read_input_tokens, cache_creation_input_tokens) and model per turn in the hook
payload, and/or persist that usage in the transcript JSONL. That would let a Stop/SessionStart hook display the session cost in the TUI via
systemMessage without patching the dist. As an alternative, an endpoint like /internal/usage/summary?sessionId= would return per-session
cost directly. Or add a cost session section in the TUI.

Would this be something you'd consider for upcoming hook events?

Use Case

No response

Additional Context

As a developer who runs long, multi-turn coding sessions with Command Code, I want to see the running cost of the current session directly
in the TUI so I can make informed decisions without leaving the terminal — for example, stopping or switching models when a session is
getting expensive, comparing cost across models (non-open vs. open models) on the same task, or tracking spend per project/client for billing.

Today the only cost visibility is the account-level /internal/usage/summary (after the fact, mixed across all sessions) or the internal
"updated" indicator, neither of which tells me what the session I'm actively in is costing. If per-turn usage and model were exposed to
hooks (or persisted in the transcript), I could write a small Stop/SessionStart hook that prints the live session cost as a systemMessage in
the feed — no dist patching, fully under my control.

How important is this to you?

Ver important.