fix(hooks): run hooks through sh on Windows when in a POSIX shell env

[timvw]

Fixes #118.

The decision

Both open questions from the issue, resolved as recommended there:

Policy — option 1 (POSIX shell only when we are demonstrably in one). On Windows, sh -c is used when isPOSIXShellEnv() (from #113) reports a POSIX environment and sh resolves on PATH; otherwise cmd /c stays. Nothing changes for PowerShell/cmd users, so anyone with %WT_PATH%-flavoured hooks is unaffected. A missing sh falls back to cmd rather than failing every hook at exec time.

The wrinkle — convert in-process, to mixed form C:/a/b, not /c/a/b. The issue is right that picking sh fixes nothing on its own: cd $WT_PATH in bash eats the backslashes as escapes. WT_PATH and WT_MAIN are converted when the hook shell is POSIX. Mixed form rather than the /c/a/b sketched in the issue, because it is strictly wider:

	C:/a/b	/c/a/b
Git Bash / MSYS2 builtins	✓	✓
Cygwin builtins	✓	✗ — drives mount at /cygdrive/c
native tool invoked by the hook (code $WT_PATH, npm)	✓	relies on MSYS argv mangling

Backslashes are what actually break the hook, so swapping the separator is the whole fix — no drive-letter special-casing, and already-POSIX paths pass through unchanged. WT_BRANCH is deliberately left alone: a branch may legitimately contain a backslash and is not a path.

No cygpath dependency, and no hooks.shell config key — not needed to close this.

Changes

• cmd/hooks.go — hookShell() picks the interpreter; adaptHookEnv() + toPOSIXPath() adapt the path-valued vars. Both take goos explicitly, matching the detectShell(args, goos) convention, and hookShell takes lookPath so the "POSIX env but no sh" branch is testable off Windows.
• cmd/hooks_test.go — new: interpreter selection across Git Bash / Cygwin / PowerShell / no-sh, path conversion, and env adaptation (including that the caller's map is not mutated).
• e2e/scenarios/hooks.yaml — the five scenarios lose skip_os: [windows] for skip_shells: [powershell, pwsh].
• docs/configuration.md — the "sh -c (or cmd /c on Windows)" footnote replaced with a table of which shell runs a hook where, the exact rule, the path form, and a cmd-syntax counterexample. docs/examples.md gets a note at the top pointing there.

Verifying

One thing the issue did not anticipate: the five scenarios also needed their assertion steps moved from $WORKTREE_ROOT to $WORKTREE_ROOT_POSIX. Those steps run test -f / cat in the harness shell, and on Windows $WORKTREE_ROOT is the native form — so they would have failed on backslashes even with the fix in place, exactly the native/POSIX split #116 introduced.

Per the issue, read the counts rather than the tick — E2E (windows, bash) should go 96 passed / 9 skipped → 101 / 4, with the PowerShell legs unchanged at 71/34 and 72/33. copy_env_file_when_exists is the scenario that actually proves the fix; its sibling copy_env_file_skipped_when_missing would pass either way.

Locally on macOS: 310 e2e passed / 0 failed / 23 skipped. go test ./... passes except TestBashTabCompletionForCheckoutBranch and TestBashTabCompletionForCommands, which fail identically on a clean main here (bash 3.2).

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 37.71%. Comparing base (4984f16) to head (afa04b2).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #120      +/-   ##
==========================================
+ Coverage   36.89%   37.71%   +0.82%     
==========================================
  Files          28       28              
  Lines        3201     3216      +15     
==========================================
+ Hits         1181     1213      +32     
+ Misses       1934     1913      -21     
- Partials       86       90       +4     

Files with missing lines	Coverage Δ
cmd/hooks.go	89.18% <100.00%> (+26.47%)	⬆️

... and 1 file with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[timvw]

CI verified — counts, not the tick

All checks green. Reading the leg summaries as #118 asks, since a scenario that silently keeps skipping is indistinguishable from a passing one in the check status:

Leg	main	this PR
E2E (windows, bash)	102 passed / 11 skipped	107 / 6	+5 / −5 ✅
E2E (windows, pwsh)	78 / 35	78 / 35	unchanged ✅
E2E (windows, powershell)	77 / 36	77 / 36	unchanged ✅

The absolute numbers differ from the 96→101 and 71/34, 72/33 predicted in the issue — that baseline predates #117 adding wt cd. The delta is exactly the +5/−5 predicted, and the PowerShell legs are untouched, which is the property that actually matters: the conservative policy did not change behaviour for anyone outside a POSIX shell env.

All five scenarios pass by name in E2E (windows, bash):

PASS: hooks/post_create_hook_runs
PASS: hooks/post_checkout_hook_runs
PASS: hooks/checkout_hooks_run_on_existing_worktree
PASS: hooks/copy_env_file_when_exists
PASS: hooks/copy_env_file_skipped_when_missing

copy_env_file_when_exists is the one that proves the fix — it asserts the .env contents actually arrived in the worktree, so it can only pass if test -f, cp, $WT_MAIN and $WT_PATH all worked under a real Git Bash. (Its sibling would have passed either way, as the issue notes.)

Unit Tests (windows) also passes, so hookShell/adaptHookEnv are exercised on a real Windows host, not just via the injected goos.