feat(bindings): add ACP thread binding adapter for WeChat (fixes #55)

[lyfuci]

Summary

Adds ACP thread binding support to the WeChat channel so
sessions_spawn(runtime="acp", mode="session", thread=true) actually works
instead of failing with thread_binding_invalid: Thread bindings are unavailable for openclaw-weixin.

Closes #55.

Why

Today every sessions_spawn(runtime="acp") from a WeChat conversation has to
fall back to mode="run" (one-shot). Each call spawns a fresh Claude Code /
Codex / Gemini CLI process, runs one turn, and loses all state:

• 10-20s cold start per turn
• No memory between turns — Claude Code can't iterate on its own work
• Multi-step tasks must be crammed into one huge prompt
• Orphan harness processes sometimes linger after acpx closes the ACP
  connection

The fix is to register a SessionBindingAdapter for the WeChat channel so
persistent ACP sessions can be tied to a WeChat 1:1 conversation, matching
what Telegram / Discord / LINE / iMessage / BlueBubbles already do.

What's in the change

Four commits, each scoped so reviewers can read them independently:

1. chore(deps): bump openclaw devDep to ^2026.4.9
   The ChannelPlugin.conversationBindings seam was added after 2026.3.23,
   so typecheck needs a newer dev dependency. Runtime host requirements are
   unchanged (still governed by openclaw.plugin.json minHostVersion).

2. test(pairing): also mock openclaw/plugin-sdk/infra-runtime subpath
   Pre-existing test regression exposed by the dev dep bump.
   pairing.ts imports withFileLock from the infra-runtime subpath, but
   pairing.test.ts only mocked the root openclaw/plugin-sdk barrel. With
   2026.4.x the subpath is a distinct module graph node so the root mock no
   longer intercepts. One-line fix: add a parallel vi.mock() for the
   subpath. Not a behavior change — just a test infrastructure fix for
   the bump.

3. test(runtime): add unit coverage for weixin runtime singleton helpers
src/runtime.ts was previously at 0% coverage. Added 6 thin unit tests
   for setWeixinRuntime / getWeixinRuntime / waitForWeixinRuntime /
   resolveWeixinChannelRuntime. This lifts baseline coverage enough to
   absorb the new file in commit 4 without tripping the 90% threshold.

4. feat(bindings): add ACP thread binding adapter for WeChat
   The actual feature:

   • src/thread-bindings.ts (~580 LOC) — per-account binding manager +
     SessionBindingAdapter + persistent JSON store + 60s idle/max-age
     sweeper. Structurally a slimmed-down port of
     extensions/telegram/src/thread-bindings.ts from the openclaw
     monorepo, with forum-topic / group handling removed.
   • src/thread-bindings.test.ts (~400 LOC, 27 tests) — bind / resolve /
     touch / unbind adapter flow, metadata merge, persistent store load /
     reload, version rejection, bad-entry rejection, sweeper expiry, sweeper
     no-op on non-expired records.
   • src/channel.ts — import createWeixinThreadBindingManager, eagerly
     initialize it in gateway.startAccount (once per account), and add the
     conversationBindings block to weixinPlugin so the openclaw ACP
     runtime discovers the adapter.

Design decisions (call out for the reviewer)

• Conversation id = <from_user_id> (e.g. xxx@im.wechat) scoped per
  bot accountId. This matches how weixinMessageToMsgContext already
  routes inbound traffic — From, To, and OriginatingTo are all set to
  from_user_id, and group_id is defined in WeixinMessage but never
  consumed anywhere in the plugin. Treating the plugin as 1:1-only keeps
  this PR small. If group / chat-room support lands later, the adapter
  should learn a child placement that maps to a room-level conversation
  id, similar to Telegram's forum topics.
• Persistent store lives under $STATE_DIR/openclaw-weixin/ thread-bindings-<accountId>.json, using the existing
  writeJsonFileAtomically helper from openclaw/plugin-sdk/json-store to
  avoid torn writes. Load errors are downgraded to logVerbose so a
  corrupt file doesn't brick the plugin.
• Default idle timeout 24h, no max age cap — identical to Telegram's
  DEFAULT_THREAD_BINDING_IDLE_TIMEOUT_MS / MAX_AGE_MS, so operators
  who already know Telegram get the same behavior.
• Only the current placement is advertised in
  SessionBindingAdapter.capabilities.placements. The bind method
  rejects placement: "child" requests outright rather than silently
  treating them as current.

Out of scope (happy to follow up)

• setIdleTimeoutBySessionKey / setMaxAgeBySessionKey —
  conversationBindings accepts these as optional hooks and Telegram wires
  them to runtime knobs. Left out to keep the PR focused; trivial to add
  once the approach here is accepted.
• Farewell-on-unbind message delivery (Telegram sends a "session closed"
  hint to the chat). WeChat delivery semantics are different enough that
  this warrants a separate design conversation.
• Group / chat-room conversation binding. Depends on the broader question
  of whether openclaw-weixin will handle group_id traffic at all.

Test plan

• npm run typecheck
• npm test — 370 tests passing, coverage 91.38% lines / 90.1% branches
  / 96.29% funcs (above the repo's 90% threshold)
• npm run build
• Live round-trip (verified separately on a local gateway): can spawn a
  persistent ACP Claude Code session from a WeChat conversation with
  mode="session", thread=true without hitting
  thread_binding_invalid, and a follow-up sessions_send is routed
  back to the same harness session.

One pre-existing unrelated note:

• src/auth/pairing.test.ts was failing on main immediately after the
  openclaw dev dep bump due to a mock subpath mismatch (see commit 2). I
  included a single-line fix for it so CI stays green. If you would rather
  land that separately I'm happy to cherry-pick it into its own PR.