fix(devlore-test): route output streams portably instead of via /dev paths - #380

…paths

Issue #379. devlore-test run could not execute on Windows at all: every output
stream defaulted to /dev/stdout, openDest was a plain os.OpenFile, and the open
failed before the command produced anything. It aborted with "opening graph
output: open /dev/null: The system cannot find the path specified."

Per the ruling, the two sentinels are handled differently because they are
different problems.

Standard output is intercepted. stdoutSentinel is matched in openDest and never
reaches the filesystem, which also fixes a latent bug the Unix path was hiding:
the callers defer a Close on what openDest returns, so routing to stdout was
closing the process's own standard output. openDest now returns an
io.WriteCloser and the stdout case is a nopWriteCloser.

Discard needs no interception. os.DevNull already names each platform's device
— /dev/null on Unix, NUL on Windows — and opens correctly on both. So the fix
there is spelling, not mechanism: help text and examples interpolate
os.DevNull rather than hardcoding /dev/null, and the same applies to the tests
that route streams away.

Four further sites carried the same defect and are corrected together, being
the identical mistake rather than separate scope: commands_test.go routed to
/dev/null directly, and scenario_integration_test.go and repo_cmd_test.go both
set GIT_CONFIG_GLOBAL and GIT_CONFIG_SYSTEM to /dev/null to suppress ambient
git configuration. No /dev/null literal remains in non-prototype Go code.

The root command's help also advertised /tmp/receipt.yaml as its example
destination; that is now a bare relative filename, which is meaningful
everywhere.

This is roughly 16 of the 47 remaining Windows failures — the 12 exit-code-1
cases, the 5 missing JSON summaries, the 2 missing "Hello World!" assertions,
and the 2 explicit /dev/null failures.

Not addressed here, and left for its own change: data/test_compensation.star
forces a failure with destination_path="/dev/null/impossible/path.txt". On
Windows that path is apparently creatable, so the copy succeeds where the
fixture expects an error. That is a different problem — the impossible-path
idiom — not the discard sentinel.

The plan's corrected phase 2 triage rides in this commit: the .exe fix cleared
16 failures rather than the ~37 estimated, because 14 TestCLI cases were
failing for this second reason behind the first.

Verified on Darwin: make vet, make build, full make test green — zero FAIL or
panic lines by count; gofmt clean; run --help renders /dev/null here, which is
what os.DevNull is on this platform. CI's windows leg is the real verification.
Plan: docs/plans/platform-test-matrix.md (phase 3).
Issue: #379