agent tests: Zed's folder on Windows is %APPDATA%\Zed, and two Codex fixtures quote a literal path

[TryWorld2026]

agent tests: Zed's folder on Windows is %APPDATA%\Zed, and two Codex fixtures quote a literal path

What is wrong

Three Windows-only test failures in internal/agent, each with its own cause.
None of them is a product bug.

1. syncHome isolated HOME and USERPROFILE but not %APPDATA% or
%LOCALAPPDATA%. An agent on Windows takes those folders from the
environment, so the package's tests read — and in places wrote — the developer's
real Roaming directory instead of the sandbox home. syncHome is the package's
shared sandbox, so this one fix covers every test that uses it.

2. TestZedRelativeConfigDirUsesDefault expected one path on every platform.
On Windows zed() takes its folder from %APPDATA% whatever cfg it is given,
so the test failed with relative config directory escaped default: <package sandbox>\AppData\Roaming\Zed\settings.json, want <home>\.config\zed\settings.json
— it was asserting magpie's behaviour was wrong on the one platform where it is
right.

3. Two Codex fixtures wrote a Windows absolute path into a TOML basic
string. In "C:\Users\…" the \T, \001 and \m are not escapes:
edit.SetTOMLTop answers line 3, column 26: invalid escaped character U+0031 '1', magpie's Codex model list is never written, and the Codex half of
TestSuffixModesReachAgentFiles fails with on, codex: map[]. The package's
other TOML fixtures (sync_test.go, modelprefs_test.go, legacy_test.go)
already quote the path as a literal.

What changed

• syncHome also sets APPDATA and LOCALAPPDATA under the sandbox home.
• TestZedRelativeConfigDirUsesDefault expects Zed's own folder under the
  APPDATA the test sets on Windows, and the cfg argument everywhere else;
  it sets APPDATA itself, as the package's other sandboxes do, so the
  expectation names its own home.
• The two Codex fixtures quote the path as a TOML literal.

Semantic change

No app change. What changes is what go test ./internal/agent exercises on a
Windows box: the Codex half of TestSuffixModesReachAgentFiles had never once
been exercised there, and the package's sandbox no longer reaches the real
%APPDATA%.

Verification

On the Windows box:

• TestZedRelativeConfigDirUsesDefault and TestSuffixModesReachAgentFiles fail
  before and pass after.
• The fixture's old form reproduces invalid escaped character U+0031 '1' from
  edit.SetTOMLTop and the literal form does not.
• go vet and GOOS=windows/darwin/linux go build -tags nogui ./internal/agent/ pass; gofmt is clean on the changed files after LF
  normalisation, since the working tree is CRLF under core.autocrlf.

Not this PR's impact

Windows CI does not run this package (the Windows job only compiles the suite),
so this does not turn a red CI job green. What it fixes is a local
go test ./internal/agent/ on Windows and the sandbox reaching the real
%APPDATA%.

A third Windows-red test in this package is not fixed here, by design.
TestCodexStaleOnlyForMagpiesChange still fails on Windows with
magpie's list changed: 0 stale, want 1, exactly as it does on main: Agent.Stale()
returns 0 on Windows unconditionally (internal/agent/connectinfo.go — it reads
ps's etime, which Windows has no way to give), so the assertion can never hold
there and the test has no platform guard. That is the neighbouring PR's change;
this one leaves it precisely as red as it already was.

Note on a neighbour

internal/agent/connect_stale_test.go is also changed by the Windows platform
guards PR, which adds a skip to TestCodexStaleOnlyForMagpiesChange where this
one only changes its TOML quoting. The two are different classes and will
conflict in the one file; whichever merges second resolves it.

Both Codex fixtures are quoted as literals because they are both the same bug,
not just the one that is reachable today: a t.TempDir() path is a Windows
absolute path on a Windows box, and in a TOML basic string its backslashes are
escapes. connect_stale_test.go's copy is currently masked by the Stale()
failure above, which reaches its assertion first; quoting it as a literal keeps
the day Stale() can answer on Windows, or the day that guard comes out, from
reopening it.