Windows builds hang when the HEAD commit message is large (Invoke-Git deadlock)

[tablackburn]

Part of #120. Found while verifying the #162 merge.

Symptom

On Windows, Set-BuildEnvironment hangs forever when the HEAD commit message is large. Nothing
is printed, no error, no timeout — the build never starts.

This is currently happening on main. The push-triggered CI run for 9720cb7 hung both Windows
legs for 27+ minutes with zero output (ubuntu and macOS passed); I cancelled it. Every local
./build.ps1 on Windows hangs the same way, because build.ps1 calls Set-BuildEnvironment -Force before Invoke-psake.

Root cause

BuildHelpers\Invoke-Git (2.0.16) redirects both output streams and then waits for the process
to exit before reading them:

$pinfo.RedirectStandardError = $RedirectStandardError # $true
$pinfo.RedirectStandardOutput = $RedirectStandardOutput # $true
...
$null = $p.Start()
$p.WaitForExit() # <-- blocks here
...
$stdout = $p.StandardOutput.ReadToEnd() # <-- never reached

When git writes more than the pipe buffer holds, git blocks waiting for someone to drain the
pipe, and Invoke-Git blocks waiting for git to exit. Neither can proceed. It is the textbook
WaitForExit-before-ReadToEnd deadlock described in the
ProcessStartInfo.RedirectStandardOutput remarks.

Get-BuildVariable calls it as Invoke-Git log --format=%B -n 1 <sha> to populate
$env:BHCommitMessage, so the payload is the full commit message body.

Reproduction

Measured on Windows 11, PowerShell 7.6.4, BuildHelpers 2.0.16, against real commits on main:

Commit	Message size	Result
f726f48	2617 bytes	returns
64b7207	5498 bytes	returns
9720cb7	5757 bytes	hangs

Import-Module BuildHelpers
Invoke-Git -Arguments 'log --format=%B -n 1 9720cb7' # never returns on Windows

Linux and macOS have a larger pipe buffer and do not hit the threshold at these sizes, which is
why only the Windows legs hang.

Why CI did not catch it before merge

On a pull_request event, GitHub checks out a synthetic merge commit whose message is short
(Merge <sha> into <sha>). The real squash message only becomes HEAD on the push to main.
Every PR will be green and then hang main if its squashed body is large enough.

This repository writes long PR bodies and squash-merges them verbatim, so the trigger is not
exotic — it is the house style. #163 and #165 both have bodies over the threshold today.

Consumer impact

This is not only a repo-internal problem. Initialize-PSBuild calls Set-BuildEnvironment, so
any consumer building on Windows hangs when their own HEAD commit message crosses the threshold.
A hang with no output is a bad failure mode to ship at 1.0.0.

Options

1. Stop passing the payload through Invoke-Git. Populate $env:BHCommitMessage ourselves
   before calling Set-BuildEnvironment, or truncate it. Narrow, immediate, ours to control
2. Fix upstream. Read the streams asynchronously (BeginOutputReadLine, or ReadToEnd
   before WaitForExit) and PR it to BuildHelpers, then raise the RequiredModules floor once
   released. Correct fix, slowest path
3. Trim squash commit messages at merge time. Pure process change, no code — but it depends
   on every future merge remembering, and it does nothing for consumers
4. Drop the BuildHelpers dependency for this variable. Largest change; worth naming only
   because [Tracking] PowerShellBuild v1.0.0 roadmap #120 already asks what 1.0.0 should depend on

Options 1 and 2 are not exclusive: 1 unblocks now, 2 fixes it for everyone.

Immediate mitigation

Until this is fixed, squash-merge with a short commit body on this repo. The full PR
description stays on the pull request either way.

Done when

./build.ps1 and Windows CI complete regardless of HEAD commit message size, with a regression
test covering a message larger than the pipe buffer.

[tablackburn]

Confirmed from both directions on main today.

Long body → hang. 9720cb7 (#162, squash body 5757 bytes): both Windows CI legs hung 27+ minutes with no output, ubuntu and macOS passed. Cancelled.

Short body → clean. 7a783e1 (#163, squash body 105 bytes): the same workflow on the same branch finished green in every leg, Windows PowerShell 5.1 included (4m). Locally, ./build.ps1 -Task Test now completes — 447 passed, 0 failed — where it hung indefinitely at the previous commit.

Same repository, same command, same machine; the only variable is the size of the HEAD commit message. That is as close to a controlled experiment as this gets, and it also means the workaround in "Immediate mitigation" is verified rather than theoretical.

One practical note for anyone hitting this: a local checkout sitting on the old commit keeps hanging until it is fast-forwarded. I lost a few minutes to exactly that after the fix was already on the remote.