test(server): isolate the keychain in the connections-route suite

[jack-champagne]

Makes unit (linux) interpretable again. It has been dying with exit code 137 and reporting nothing.

Not an OOM

Turbo labels any signal-killed child "likely due to running out of memory." That label is boilerplate and it is wrong here. Bun's own report:

Elapsed: 1498ms | RSS: 0.15GB | Peak: 0.37GB | Faults: 0 | Machine: 32.38GB
panic(main thread): Segmentation fault at address 0x0

0.37 GB peak on a 32 GB machine, zero faults, dead in 1.5 s. A larger runner would not have helped (and is unavailable — the org is plan=free; GET /orgs/harmoniqs/actions/hosted-runners returns 404).

What it actually was

One test — amicode-connections-routes.test.ts:335, "full lifecycle: submit(valid) → …" — segfaults at connections.ts:1005:

pasqalSecretStore().write(PASQAL_SECRET_ACCOUNT, { username, password })

which reaches @napi-rs/keyring's native Entry.setPassword(). Traced with instrumentation: [T3] before keychain prints, [T4] after keychain write never does.

That write sits only on the valid && token !== null branch, which explains the pass/fail pattern exactly:

test	branch	keychain write	result
full lifecycle (valid + token)	valid + token	yes	crash
null token → session-only	valid, null token	no	pass
failure classes (exit 2/4/1)	config / invalid	no	pass

Bisected to 467fb15cf ("two-step Pasqal auth", #194), which introduced that write per the ADR 0001 addendum. The test passed at 3e4b37298, the commit immediately before.

Production is not affected

Verified end-to-end, not assumed. Real Pasqal login, real pasqal_validate.py, real pasqal-cloud SDK, real keychain, against bun run … serve:

POST /amicode/connections/credential
  ok=True  state=connected  devices=[{'name':'FRESNEL'},{'name':'FRESNEL_CAN1'}]
server: ALIVE

And it demonstrably reached the crash site — the keychain slot was written and the token persisted (token_len: 1278).

context	keychain write runs	result
bun test + Effect route	yes	segfault
bun run … serve + Effect route	yes	survives
bun test, keyring alone	yes	survives
bun run, keyring alone	yes	survives

Only the first combination crashes. It is a Bun-test-runner/NAPI interaction, not a product fault, and 467fb15cf is not shipping a crash to users.

The fix

Install the in-memory secret store in beforeEach — exactly what the sibling amicode-connections.test.ts already does, which is why that file passes:

restoreSecretStore = setPasqalSecretStore(inMemorySecretStore())

Also adds AMICO_PASQAL_KEYCHAIN_SERVICE to ENV_KEYS. pasqal-secret.ts documents it as existing so "an isolated sandbox / test run gets its own slot and never shares the real connection's secret"; it was unused here. Belt and braces, so a leak cannot reach the real slot even if the seam is ever removed.

Worth noting independently of the crash: this suite was driving the developer's real OS keyring.

Stub validator: bun → python3

stageStubValidator() set AMICO_PYTHON = process.execPath, i.e. bun standing in as the Python interpreter. In production amicode provisions <opsDir>/venvs/pasqal-connector and passes it as AMICO_PYTHON (harmoniqs/amicode#189, the bundled-Julia pattern), so the interpreter is always a real python. The old stub tested a configuration that never ships.

Now a python3-executed .py with the same record/scenario contract, plus actions/setup-python in the unit job. The stub needs an interpreter, not the SDK — integration against the real pasqal-cloud SDK stays where it belongs, in amicode's assert_provisioned_python.mjs gate.

The env assertion now filters interpreter-owned locale vars, because CPython injects LC_CTYPE into its own environ under PEP 538 C-locale coercion. It still asserts the real thing: that the spawner declared nothing beyond PASQAL_USERNAME/PASSWORD/PROJECT_ID + PATH.

Windows unit lane dropped

opencode.lock.json ships darwin-arm64 and linux-x64 only, so a windows unit lane gates a platform this fork does not distribute — and at ~50 min it is the long pole on every run. It was also red on arrival for a reason unrelated to anything under test: POSIX path separators hardcoded in amicode-vaults.test.ts (#76). e2e still runs both platforms.

Verification

amicode-connections-routes.test.ts     13 pass / 0 fail      (was: segfault)
test/server/  (63 files)              554 pass / 2 skip / 0 fail
bun turbo typecheck                    23 successful, 23 total

Before this, the full packages/opencode suite could not complete at all. It now runs 3234 tests to completion, leaving a readable list: the stale LLM fixtures (#80) and a umask-dependent permissions assertion in tool.write (0o644 vs 0o664 — passes on CI's umask 0022, fails on a developer umask 0002).

Refs #82, #76