ci(openclaw): run the plugin unit suite per PR and the E2E suite on staging/main

[harrymove-ctrl]

The OpenClaw plugin had tests that ran nowhere, and E2E coverage that lived in a gitignored scratch directory.

#668 added test/security.test.mjs for the GH #639 / #640 injection work and #662 added test/plugin.test.mjs, but oc-memwal appears nowhere in test.yml, so both only ran if someone typed the pnpm filter by hand. The regression tests guarding a security fix were not guarding anything.

Two jobs, split by cost

OpenClaw Plugin / Unit tests — every PR. Runs test/*.test.mjs: 30 cases, no credentials, no network.

OpenClaw Plugin / E2E — pushes to staging and main only, never per PR. Runs test/e2e/*.test.mjs.

The split exists because Walrus storage is append-only with no per-blob delete, so every E2E write is permanent and costs gas plus storage. Running that per PR would be wrong.

The E2E suites

test/e2e/mock-relayer.test.mjs serves a deliberately broken relayer from the test process — rate limiting, 5xx, and a socket held open with no reply. No credentials, no network, no cost. The hang case is the point: before the request deadline existed, a relayer that accepted the connection and went quiet left the recall hook pending forever and the turn never completed. An unreachable hostname does not reproduce that, because DNS fails fast.

test/e2e/live-relayer.test.mjs runs against a real relayer and skips itself unless credentials are present, so it is safe locally and in any job without secrets. Writes are separately opt-in behind MEMWAL_E2E_WRITE=1 and land in a throwaway e2e-<timestamp> namespace, never default. Authorisation is proven with a signed recall() rather than /health, which is unauthenticated and answers even for a revoked key.

Secrets

The E2E job reads MEMWAL_E2E_PRIVATE_KEY and MEMWAL_E2E_ACCOUNT_ID. Neither exists yet, and that is fine — without them the live cases skip and the mock cases still run, so this job cannot fail a branch. It targets the staging relayer even on a main merge, since there is no reason to leave test data on mainnet.

What the cases hold the line on

Each maps to a defect found by running the plugin against a real gateway:

• the manifest must not mark credential fields required, or plugins install deadlocks in both directions
• contracts.tools must stay declared, or neither agent tool registers and the documented tools.allow step is impossible
• relayer calls must keep their deadline, or a stalled relayer holds the turn open
• the injection filter must not regress either way — attack payloads stay caught, and ordinary developer speech stays storable, which is the half that breaks silently

Verification

• unit: 30 tests, 30 pass
• e2e without credentials: 10 tests, 4 pass, 6 skipped, 0 fail — the live half skips cleanly
• on this PR the E2E job correctly reports skipping

One detail worth flagging for review: the hang case initially took 301s despite the deadline firing at 1.5s. withTimeout races the request rather than aborting it, because recall() builds its own AbortController and accepts no external signal, so the abandoned fetch kept its socket and server.close() waited on it. The mock now force-closes connections. That socket leak is real in production too; fixing it properly means threading an AbortSignal through the SDK, which is out of scope here.