fix(env): translate PEBREL_HOOK_EXE across the WSL boundary

[AnxForever]

Result / 用户结果

PEBREL_HOOK_EXE, and its legacy alias NEBULA_HOOK_EXE, now reach WSL panes as a /mnt/... path. They are translated the same way as PEBREL_CLI and PEBREL_BIN_DIR.

Before this change the variable did not cross WSLENV at all. Inside a WSL pane:

$ echo "hook=[$PEBREL_HOOK_EXE]"
hook=[]

The only value a user could find was the process-wide Windows one, D:/tools/Pebrel/runtime/pebrel-hook.exe, and Linux cannot execute that. A Claude Code hook wired with it on the WSL side fails on every event:

SessionStart:startup hook error
Failed with non-blocking status code: Error occurred while executing hook command:
ENOENT: no such file or directory, posix_spawn 'D:/tools/Pebrel/runtime/pebrel-hook.exe'

Only SessionStart shows the error. PostToolUse, UserPromptSubmit, Stop and the other events fail silently. The same hooks work when the path is /mnt/d/tools/Pebrel/runtime/pebrel-hook.exe, which is the value this change exports.

Design / 设计边界

• Responsibility: agent_env.rs is the one place that decides what crosses into WSL through WSLENV. This adds PEBREL_HOOK_EXE/p and NEBULA_HOOK_EXE/p to WSLENV_ENTRIES, next to PEBREL_CLI/p and PEBREL_BIN_DIR/p.
• Why here: the doc comment above the table already describes this failure: without /p, the guest shell gets a D:\… literal that nothing on the Linux side can execute. HOOK_EXE_ENV exists so that plugins can shell out to the hook bridge. It has the same process-wide scope as the other two paths (ai_hook.rs), so it belongs in the same table.
• Unchanged: hook delivery itself. Pane identity already crosses through WSLENV, and the notify pipe is reachable from WSL. Injecting a hook payload from a WSL pane flips pebrel agent list to state_source: hook. The exported path was the only missing piece.
• Compatibility: no interface, data-format or threading change. WSLENV_ENTRIES is #[cfg(windows)], so other platforms are unaffected, and the existing merge still preserves entries that are already in WSLENV.
• Docs: docs/runtime-control-api.md lists the translated entries; this adds the hook path to that list.

Evidence / 验证依据

All checks ran on Windows x64 with the pinned toolchain (Rust 1.97.1):

• cargo test --locked -p nebula --features gpui-test-support wslenv_entries_match_variables: passes.
• Regression check: I removed the two table entries, kept the new assertion, and ran the same test. It fails:

  test agent_env::tests::wsl::wslenv_entries_match_variables ... FAILED
  PEBREL_HOOK_EXE missing from WSLENV passthrough: ["PEBREL_PANE_ID", "NEBULA_PANE_ID", "TERM_PROGRAM", "TERM_PROGRAM_VERSION", "COLORTERM", "PEBREL_CLI", "PEBREL_BIN_DIR", "PEBREL_PROCESS_ID", "PEBREL_RUNTIME_ENDPOINT", "NEBULA_CLI", "NEBULA_BIN_DIR"]
  

• Full cargo test --locked -p nebula --features gpui-test-support, run on a v1.9.1-based branch that carries this commit: 2024 passed, 0 failed.
• cargo fmt --all -- --check: clean. I ran it with --config newline_style=Auto because this Windows checkout uses CRLF.
• python3 scripts/check_architecture.py --base origin/main: exit 0.
• Not run: Linux and macOS, since the change is #[cfg(windows)]. I also haven't done an end-to-end check in a live WSL pane with a build of this exact branch.

Required Review / 必须确认

• I followed CONTRIBUTING.md, docs/architecture.md, and docs/project-constraints.md.
• I split responsibilities, not arbitrary line ranges; no duplicate behavior authority was added.
• python3 scripts/check_architecture.py --base <PR-base-commit> passes; budgets were not inflated to fit the change.
• Tests cover success and failure; platform/feature coverage limitations are stated.