test(simulator): prove the OpenClaw plugin acquires its own client

[chughtapan]

Closes the last three open items from the slate's verification table.

1. The decisive assertion now runs

The plan's step-10 gate ends with: harness-adapters.integration.test.ts → startOpenClawGateway constructs the plugin as createMoltzapChannelPlugin() with no injected client and still completes the round trip. The OpenClaw case injected one, so that assertion had never executed — the only end-to-end proof of the production acquisition path was the NanoClaw case.

The plugin now acquires the slot's client itself. One slot names one loopback port, so nothing else may open a second daemon against it: the peer opens the conversation and the membership boundary is asserted from the peer's side, which is the shape runNanoClawCase already had. assertConversationBoundary is replaced by assertPeerConversationBoundary.

Coverage moved, not lost: HarnessClientService.startConversation is still exercised by moltzapd.integration.test.ts → starts a conversation and round-trips a bound reply through MCP only and by harness-client.test.ts.

2. harnessClientForAccount stays — a recommendation I reversed

I previously said this should be deleted, and step 10's grep does list it. Having looked at what depends on it, deleting it is the wrong call:

• It returns HarnessClientService — the sanctioned adapter-facing capability. It is not a second route to the daemon the way the deleted createService / createCore were, which returned MoltZapService / MoltZapChannelCore.
• Four OpenClaw suites drive pure presentation logic (inbound projection, delivery formatting, contract properties) through a fake turn stream. Deleting the seam converts all four into daemon-spawning integration tests, for no safety gain.

So the field stays, marked @internal and documented as test-only in both the plugin type and the fixture header. git grep harnessClientForAccount -- packages/ | grep -v '\.test\.ts\|test-utils/' returns only the declaration and its own read — no shipped source supplies it, and #962's checkAdapterFile rule already blocks the real bypasses by symbol and subpath.

I'd rather flag this as a plan step that stopped making sense than execute it and degrade the test suite.

3. Divergence ledger — rows 1 and 6 closed

Row 1 was the corpus's only confirmed binding divergence and was still open: packages/evals had neither AGENTS.md nor CLAUDE.md, packages/simulator had no CLAUDE.md. All seven packages now carry both.

Row 6 — ConversationWithParticipants has zero consumers outside packages/client, but it names the return type of HarnessClientService.startConversation, so the export is correct and the doc comment was the wrong half. Rewritten to say what is true.

Rows 2 and 3 remain: contested, v2-owned, unresolvable until a main-resident HarnessClient contract is admitted. Row 5 was withdrawn (v2 renamed on its own branch; main's table matches main's tree). Row 7 closed in #961.

Full re-derived ledger and the per-row verification table are posted to #926: #926 (comment)

Verification

Run against the new gate shape, since workspace:precommit no longer exists after main's hook retiering:

• pnpm typecheck — 0
• pnpm lint — 0 (includes arch:check across 9 projects and the new lint:script-reachability)
• pnpm format:check — 0
• pnpm docs:generate — no drift
• simulator integration suite — 3/3, OpenClaw 59s, NanoClaw 45s, restart 46s

One miss worth recording: deleting assertConversationBoundary orphaned Schema, agentName, and INITIAL_CONTENT. The integration test passed regardless — only pnpm lint catches that, and I reported the item done before running it.

[chughtapan]

Superseded by #974, which collapses this stack into one PR against main. This branch's commits are all contained in it — nothing is lost, and this thread stays readable. Closing to leave one place to review.