fix: strip redundant thinking blocks from visible WS stream replies (#187)

[singleGanghood]

Summary

Fixes #187 — model reasoning/thinking content leaked to WeCom users in WS streaming replies, and visible text was duplicated.

Root cause

OpenClaw can deliver thinking content through two independent channels:

1. A dedicated reasoning stream (onReasoningStream → state.reasoningText)
2. <think>...</think> blocks echoed into the visible text stream (deliver → state.accumulatedText)

When both channels carried thinking content (a common pattern with reasoning-capable models), buildWsStreamContent produced nested <think> tags like:

<think>reasoning stream content</think>
<think>visible-stream thinking block</think>
visible text

The WeCom client then surfaced the inner thinking block as visible text and duplicated the visible reply.

Changes

wecom/ws-monitor.js

• Added resolveStreamVisibleText(accumulatedText, reasoningText):
  • When a reasoning stream is present, <think> blocks echoed into the visible text stream are stripped (they are redundant — the reasoning is already reported via the reasoning stream).
  • When the model writes thinking directly into the text stream (no reasoning stream), the blocks are preserved and normalized so the WeCom client can render the official collapsed "已完成思考" block (existing behavior, covered by ws.e2e.test.js).
• Applied it in all forwarding paths: finishThinkingStream, sendReasoningUpdate, sendVisibleUpdate, rotateStream, and the passive markdown candidate.

wecom/callback-inbound.js

• resolveCallbackFinalText now parses out thinking blocks (via parseThinkingContent) before normalizing the final reply text, so thinking content never reaches users through the callback path either.

Tests

• New tests/ws-monitor.think-leak.test.js:
  • Strips redundant thinking blocks from the visible reply when a reasoning stream exists (regression for Model reasoning (thinking) content leaks to users in WS streaming replies; visible text duplicated #187).
  • Preserves thinking blocks when the model writes them directly into the visible stream (official dynamic thinking stream behavior).
• Updated tests/callback-inbound.test.js to assert thinking blocks are stripped from the final callback text.
• Full suite: 321 passing (baseline 316 + 5 new tests). The only 2 failures are pre-existing macOS sandbox path issues (/private/var symlink) unrelated to this change.

Verification

• npm test: 321 pass / 2 fail (2 pre-existing environment failures)
• The official dynamic thinking stream e2e test (matches the official dynamic thinking stream behavior) still passes.
• Confirmed the new leak regression test fails on the original code and passes with this fix.