fix(runner): survive a missing run scratch dir when sourcing a test file

[Chemaclass]

🤔 Background

Related #1163

load_test_files sources each file as source "$file" 2>"$dir/source_err".
When the run's scratch directory is missing the redirect fails, and bash
reports that as the command failing — exit 1 with nothing written to the
capture file. The runner read that as the file failing to source:

✗ Error: Source
Failed to source 'b_test.sh' (exit 1, 52 bytes, no stderr)

against a file that was complete and valid. The message names the test file,
which is the one thing not at fault.

💡 Changes

• Restore the directory before sourcing, falling back to /dev/null if that is refused — losing one file's stderr capture beats failing the file for a reason that is not its own
• Acceptance test covering both halves: a run survives the directory disappearing, and a genuine source failure is still reported

Note on #1137

Not a fix for it. That flake did not reproduce in twelve isolated Bash 3.0
runs — six with this change and six without — so whether this mechanism is what
hits CI there is unproven. My earlier reproduction used a harness that shared
one working copy across containers, and I have retracted it on the issue.

[Chemaclass]

🤔 Background

Related #1163

load_test_files sources each file as source "$file" 2>"$dir/source_err".
When the run's scratch directory is missing the redirect fails, and bash
reports that as the command failing — exit 1 with nothing written to the
capture file. The runner read that as the file failing to source:

✗ Error: Source
Failed to source 'b_test.sh' (exit 1, 52 bytes, no stderr)

against a file that was complete and valid. The message names the test file,
which is the one thing not at fault.

💡 Changes

• Restore the directory before sourcing, falling back to /dev/null if that is refused — losing one file's stderr capture beats failing the file for a reason that is not its own
• Acceptance test covering both halves: a run survives the directory disappearing, and a genuine source failure is still reported

Note on #1137

Not a fix for it. That flake did not reproduce in twelve isolated Bash 3.0
runs — six with this change and six without — so whether this mechanism is what
hits CI there is unproven. My earlier reproduction used a harness that shared
one working copy across containers, and I have retracted it on the issue.

[Chemaclass]

🤔 Background

Related #1163

load_test_files sources each file as source "$file" 2>"$dir/source_err".
When the run's scratch directory is missing the redirect fails, and bash
reports that as the command failing — exit 1 with nothing written to the
capture file. The runner read that as the file failing to source:

✗ Error: Source
Failed to source 'b_test.sh' (exit 1, 52 bytes, no stderr)

against a file that was complete and valid. The message names the test file,
which is the one thing not at fault.

💡 Changes

• Restore the directory before sourcing, falling back to /dev/null if that is refused — losing one file's stderr capture beats failing the file for a reason that is not its own
• Acceptance test covering both halves: a run survives the directory disappearing, and a genuine source failure is still reported

Note on #1137

Not a fix for it. That flake did not reproduce in twelve isolated Bash 3.0
runs — six with this change and six without — so whether this mechanism is what
hits CI there is unproven. My earlier reproduction used a harness that shared
one working copy across containers, and I have retracted it on the issue.

[Chemaclass]

🤔 Background

Related #1163

load_test_files sources each file as source "$file" 2>"$dir/source_err".
When the run's scratch directory is missing the redirect fails, and bash
reports that as the command failing — exit 1 with nothing written to the
capture file. The runner read that as the file failing to source:

✗ Error: Source
Failed to source 'b_test.sh' (exit 1, 52 bytes, no stderr)

against a file that was complete and valid. The message names the test file,
which is the one thing not at fault.

💡 Changes

• Restore the directory before sourcing, falling back to /dev/null if that is refused — losing one file's stderr capture beats failing the file for a reason that is not its own
• Acceptance test covering both halves: a run survives the directory disappearing, and a genuine source failure is still reported

Note on #1137

Not a fix for it. That flake did not reproduce in twelve isolated Bash 3.0
runs — six with this change and six without — so whether this mechanism is what
hits CI there is unproven. My earlier reproduction used a harness that shared
one working copy across containers, and I have retracted it on the issue.

[Chemaclass]

🤔 Background

Related #1163

load_test_files sources each file as source "$file" 2>"$dir/source_err".
When the run's scratch directory is missing the redirect fails, and bash
reports that as the command failing — exit 1 with nothing written to the
capture file. The runner read that as the file failing to source:

✗ Error: Source
Failed to source 'b_test.sh' (exit 1, 52 bytes, no stderr)

against a file that was complete and valid. The message names the test file,
which is the one thing not at fault.

💡 Changes

• Restore the directory before sourcing, falling back to /dev/null if that is refused — losing one file's stderr capture beats failing the file for a reason that is not its own
• Acceptance test covering both halves: a run survives the directory disappearing, and a genuine source failure is still reported

Note on #1137

Not a fix for it. That flake did not reproduce in twelve isolated Bash 3.0
runs — six with this change and six without — so whether this mechanism is what
hits CI there is unproven. My earlier reproduction used a harness that shared
one working copy across containers, and I have retracted it on the issue.

[Chemaclass]

🤔 Background

Related #1163

load_test_files sources each file as source "$file" 2>"$dir/source_err".
When the run's scratch directory is missing the redirect fails, and bash
reports that as the command failing — exit 1 with nothing written to the
capture file. The runner read that as the file failing to source:

✗ Error: Source
Failed to source 'b_test.sh' (exit 1, 52 bytes, no stderr)

against a file that was complete and valid. The message names the test file,
which is the one thing not at fault.

💡 Changes

• Restore the directory before sourcing, falling back to /dev/null if that is refused — losing one file's stderr capture beats failing the file for a reason that is not its own
• Acceptance test covering both halves: a run survives the directory disappearing, and a genuine source failure is still reported

Note on #1137

Not a fix for it. That flake did not reproduce in twelve isolated Bash 3.0
runs — six with this change and six without — so whether this mechanism is what
hits CI there is unproven. My earlier reproduction used a harness that shared
one working copy across containers, and I have retracted it on the issue.

[Chemaclass]

🤔 Background

Related #1163

load_test_files sources each file as source "$file" 2>"$dir/source_err".
When the run's scratch directory is missing the redirect fails, and bash
reports that as the command failing — exit 1 with nothing written to the
capture file. The runner read that as the file failing to source:

✗ Error: Source
Failed to source 'b_test.sh' (exit 1, 52 bytes, no stderr)

against a file that was complete and valid. The message names the test file,
which is the one thing not at fault.

💡 Changes

• Restore the directory before sourcing, falling back to /dev/null if that is refused — losing one file's stderr capture beats failing the file for a reason that is not its own
• Acceptance test covering both halves: a run survives the directory disappearing, and a genuine source failure is still reported

Note on #1137

Not a fix for it. That flake did not reproduce in twelve isolated Bash 3.0
runs — six with this change and six without — so whether this mechanism is what
hits CI there is unproven. My earlier reproduction used a harness that shared
one working copy across containers, and I have retracted it on the issue.

[Chemaclass]

🤔 Background

Related #1163

load_test_files sources each file as source "$file" 2>"$dir/source_err".
When the run's scratch directory is missing the redirect fails, and bash
reports that as the command failing — exit 1 with nothing written to the
capture file. The runner read that as the file failing to source:

✗ Error: Source
Failed to source 'b_test.sh' (exit 1, 52 bytes, no stderr)

against a file that was complete and valid. The message names the test file,
which is the one thing not at fault.

💡 Changes

• Restore the directory before sourcing, falling back to /dev/null if that is refused — losing one file's stderr capture beats failing the file for a reason that is not its own
• Acceptance test covering both halves: a run survives the directory disappearing, and a genuine source failure is still reported

Note on #1137

Not a fix for it. That flake did not reproduce in twelve isolated Bash 3.0
runs — six with this change and six without — so whether this mechanism is what
hits CI there is unproven. My earlier reproduction used a harness that shared
one working copy across containers, and I have retracted it on the issue.