Feature: official Grok CLI (xAI) support

[Sallvainian]

Describe your idea
Official support for xAI's Grok CLI (grok) as a built-in adapter profile, alongside claude, codex, gemini, copilot and antigravity.

Why is this needed?
Grok can drive bmad-loop today only through a hand-written project profile, and that path has gaps:

• Grok runs each subagent as its own session and fires SessionEnd when one finishes. bmad-loop scores any SessionEnd as a crash, so a profile that maps it has every session that outlives a subagent killed mid-work. Leaving SessionEnd unmapped (as codex does) avoids that, but real crashes are then only caught by the window-death fallback.
• No usage parser reads grok's transcripts, so token budget warnings and limits never apply to grok stages.
• Grok's SessionStart names no transcript path, so the budget guard can't sample usage until the first turn ends.

Built-in support would let grok users run the loop with bmad-loop init --cli grok and nothing hand-maintained, with the same crash detection and budget enforcement the other CLIs get.

How should it work?

• A built-in grok profile: Claude-shaped hooks in .grok/hooks/bmad-loop.json, skills in .agents/skills (where the BMad installer puts grok's), --always-approve for unattended runs, and a /hooks-trust first-run note.
• Ignore a SessionEnd whose payload carries grok's subagentType marker. The main session's SessionEnd lacks it, and grok's hook docs say the field exists so a host can tell a child's teardown from its own. No other CLI sends it, and an unmarked SessionEnd still ends the session.
• A grok-updates usage parser that sums each turn's turn_completed usage in updates.jsonl, including subagent sessions.
• A transcript_template profile key that derives the transcript path from the session id and cwd at SessionStart.

PR
I understand grok isn't officially supported at the moment, so I haven't opened a PR. However, a ready implementation is on my fork, to be opened as a PR on request:

• Branch: https://github.com/Sallvainian/bmad-loop/tree/feature/grok-profile (4 commits on current main, about 600 lines with tests)
• Diff: main...Sallvainian:bmad-loop:feature/grok-profile

Additional context

• The subagentType marker was confirmed against live grok 1.0.41 hook payloads; the profile's floor is 1.0.41 for that reason.
• The parser's totals matched independent recounts on two unattended loop runs (grok 1.0.40).
• On the fork's CI, the test, lint, typecheck and build jobs pass on Linux and Windows across Python 3.11–3.14.