[enowdev]

go test ./... fails on every checkout of main, and only on CI. Two independent causes, both fixed here; no test is skipped, weakened or deleted.

1. The migration fixtures were never committed

.gitignore had an unanchored AGENTS.md line under "local agent/editor tooling". Unanchored, it also matched the fake homes of other agents under internal/migrate/testdata — nanobot, OpenClaw, OpenHuman and ZeroClaw all keep their own AGENTS.md, and OpenHuman's memory/memory.db fell under *.db.

Evidence:

$ git check-ignore -v internal/migrate/testdata/nanobot/home/.nanobot/workspace/AGENTS.md
.gitignore:35:AGENTS.md	internal/migrate/testdata/nanobot/home/.nanobot/workspace/AGENTS.md
$ git ls-tree -r --name-only 288416f | grep -E 'testdata/.*(AGENTS|CLAUDE\.local)\.md$'
(no output — never tracked in any commit; no dangling blob in the object store either)

So the files existed only on the machine that wrote the goldens, and TestNanobotPlanGolden, TestNanobotConflicts, TestNanoclawDetect, TestNanoclawPlanV2, TestOpenClawPlan, TestOpenhumanDetect, TestOpenhumanPlan, TestZeroclawDetect and TestZeroclawPlan failed everywhere else.

The fix anchors the rule to /AGENTS.md (repo root only) and un-ignores the fixture .db, then commits the fixtures themselves.

About the golden updates

Four goldens change, and every changed line is a fingerprint of the restored fixture bytes:

golden	changed line
nanobot/plan.golden	agents_md sha=
openclaw_plan.txt	agents_md sha=, role:work prompt_sha=
openhuman/plan.golden	agents_md sha=
zeroclaw/plan_v3.golden	agents_md sha=, role:coder prompt_sha=

The original bytes are not recoverable — they were never in git history and no dangling blob holds them, so the fixtures are reconstructed. What is recoverable is exact: each golden records the trimmed byte length of the text it rendered (113 / 42 (x2) / 80 chars) and the NanoClaw memory entry's own text, so the reconstructions hit those lengths to the byte. Only ContentHash output moved; every status, reason, count, order and summary line in those goldens is byte-identical, and the OpenHuman memory.db warning and both role items are preserved (the first attempt at -update silently dropped them, which is how this was caught). Run with -update twice and the diff is empty.

2. config.Load() served a config from the wrong home

The cache ignored which home it came from, so tests that set ANTARES_HOME per test got the previous test's config:

$ go test ./cmd/antares -run 'TestDaemonStateRoundTripAndStaleCleanup|TestDevicePairCLI'
--- FAIL: TestDevicePairCLI
    device_test.go:44: url without a daemon should come from config: "http://127.0.0.1:1"

TestDevicePairCLI sets ANTARES_PORT=18999, but Load() returned the config cached for the previous test's home, whose config.yaml pins port 1. Run alone, it passes. This is a real bug beyond tests: anything that changes ANTARES_HOME/ANTARES_CONFIG in-process got another directory's settings. The cache now records the home and config path it was loaded from and only serves itself while both still match.

Verification

• go vet ./... clean, go test ./... green (previously 10 failures)
• git status --ignored under testdata is empty — nothing else is silently missing
• Fixture lengths verified against every golden's documented · N chars

CI on this PR is the real check: the smoke job has been skipped on main for weeks because it needs go and web, so it runs again here.