fix(fleet-data/scrape): generalize model-label sniff to claude / gemini / o-series

[NagyVikt]

Summary

fleet-data::scrape::extract_model_label previously called line.find(\"gpt-\"), so it only surfaced a label for codex panes. fleet-launcher also spawns claude, gemini, and claw panes — those rendered with no model label in the artboard's WORKING ON column.

This PR replaces the single-prefix search with a small const slice of model-family prefixes and picks the prefix with the earliest occurrence. The rest of the trailing-word logic is identical.

Prefixes now supported

const MODEL_PREFIXES: &[&str] = &[
    \"gpt-\", \"claude-\", \"gemini-\", \"o1-\", \"o3-\",
    \"sonnet-\", \"opus-\", \"haiku-\",
];

Tiebreak rule

When multiple families appear in a single line, the one at the earliest byte index wins. Rationale: the pane's actual status line is rendered at a fixed position; later occurrences usually come from scrollback or prompt text mentioning another model. (Longest-match was considered but discarded — claude- and gemini- would never tie with each other, and the natural reading order is left-to-right.)

Tests

Added 4 new test cases covering claude-, gemini-, o3-, and the earliest-prefix tiebreak. All 70 fleet-data tests pass (66 existing + 4 new).

Test plan

• cargo test -p fleet-data — 70 passed, 0 failed
• cargo check -p fleet-state -p fleet-waves -p fleet-watcher — all clean
• Existing gpt- tests unchanged and still passing