[FEATURE] Add model, stop_reason, token usage and tool timing to hook payloads

### Preflight Checklist

- [x] I have searched [existing requests](https://github.com/anthropics/claude-code/issues?q=is%3Aissue%20label%3Aenhancement) and this feature hasn't been requested yet
- [x] This is a single feature request (not multiple features)

### Problem Statement

I maintain an open-source desk device + plugin family ([Rubato_Plugins](https://github.com/lovaxi/Rubato_Plugins)) that mirrors agent state on physical hardware for several hosts (opencode, Cursor, OpenClaw, Codex). For Claude Code I use the official hooks — seven of them, from SessionStart to StopFailure.

The problem: **hook payloads carry almost no data** (session_id, transcript_path, cwd). To know the model, stop_reason, token usage, or tool names, my plugin must tail the session transcript JSONL in parallel — a second file-watcher per session, racing with Claude's own writes, parsing a format that is not documented as a stable contract.

Consequences:

- every hook consumer duplicates the same transcript-parsing layer (fragile, version-sensitive);
- extra latency between what Claude is doing and what external surfaces can display;
- streaming metrics (TTFT, tokens/sec) are simply unavailable at any latency — so honest duration estimation ("how long will this run take?") is impossible for external hardware and status surfaces, while the data itself exists inside the session.

### Proposed Solution

Keep the hooks exactly as they are — just make the payload self-sufficient, additively:

- on **all hooks**: include `model`
- on **Stop / StopFailure**: include `stop_reason`, `duration_ms`, and `token_usage { input, output, cache_read, cache_creation }`
- on **PostToolUse**: include `tool_duration_ms`
- (stretch) a turn-level metrics field with TTFT / output tokens-per-second — I understand streaming metrics may not be feasible; the usage fields alone are the high-value part.

Ideal UX: a hook consumer can mirror the full session — state, model, cost, honest durations — from hook events alone, without opening a single transcript file. All new fields are additive: existing consumers that ignore them are unaffected.

### Alternative Solutions

- Current workaround: tail `~/.claude/projects/.../*.jsonl` from a parallel daemon — works, but it parses an internal format, races with Claude's own writes, and still cannot see TTFT or streaming rate.
- The statusline input carries some session context, but it is a render surface, not an event stream an external device can subscribe to.
- An MCP server is the wrong layer for this: MCP servers serve the model; they do not observe the session.

### Priority

High - Significant impact on productivity

### Feature Category

Developer tools/SDK

### Use Case Example

1. I kick off a long refactor with Claude Code; my desk device shows Thinking → Generating so I stop alt-tabbing into the terminal.
2. On Stop, I want the device to show "took 4m 12s, 18k tokens" — so the user can honestly decide whether the next run justifies staying at the desk, and whether a break is due.
3. Today the Stop hook hands me a transcript path; a parallel daemon tails the JSONL, parses internal message shapes, and still cannot compute TTFT or per-request usage cleanly.
4. With structured hook payloads, the same integration shrinks to ~100 lines of version-stable code — and every status bar, dashboard, and device in the hooks ecosystem gets it for free.

### Additional Context

- The integration is shipped and verified end-to-end: [Rubato_Plugins](https://github.com/lovaxi/Rubato_Plugins) (seven hooks in use, including StopFailure and idle_prompt recovery).
- Cross-host parity: opencode events already carry per-call token usage, and OpenAI Codex exposes usage in its session data — Claude Code hooks are the gap.
- My device wire contract reserves `estSec` (estimated seconds); structured usage in hooks is the missing input for an honest estimate. Strictly additive proposal — no breaking changes.

Start by finding where Claude Code builds official hook payloads for SessionStart through StopFailure, especially Stop, StopFailure, and PostToolUse. Compare those payloads with the transcript JSONL data under ~/.claude/projects/.../*.jsonl to see where model, stop_reason, token_usage, duration_ms, and tool_duration_ms are already available. Done means the listed fields are added additively to hook events without requiring consumers to read the transcript.

Vom Indexierungsmodell aus dem Issue-Text verfasst.