feat(qwen): add first-class Cognee memory extension

[netbrah]

Summary

• add a self-contained Qwen Code extension with Gemini-shaped manifests and Qwen's Claude-style hook events
• capture supported interactive prompts, tool traces, assistant responses, PreCompact anchors, and SessionEnd graph sync through the Cognee HTTP backend
• isolate Qwen runtime state and backend selection under ~/.cognee-plugin/qwen/ and COGNEE_QWEN_BACKEND
• use Qwen's millisecond timeout units, external-hook root contract, and injected skill-base path semantics
• add shared-suite, packaging, host-semantics, HTTP compatibility, Linux, and Windows coverage
• document the current Qwen prompt-provenance boundary and upstream prerequisite feat(hooks): expose send type on UserPromptSubmit payloads QwenLM/qwen-code#9511

Host-specific hardening

• PreCompact emits hookSpecificOutput.additionalContext, which Qwen actually consumes
• prompt capture and recall use submitted_prompt so ToolResult/Hook continuations cannot replace the real user question
• installed memory skills resolve extension scripts from Qwen's injected skill base instead of an undeclared Claude variable
• external hook files require ${CLAUDE_PLUGIN_ROOT}; ${extensionPath} is rejected for this loader path

Qwen currently omits submitted_prompt for ACP, headless, serve, SDK, remote-input, restored, and Vim-mode inputs while also omitting send type from command-hook payloads. Those model-bound prompts are intentionally skipped rather than guessed from prompt; tool/assistant/PreCompact/session hooks remain active. The additive host field needed to support those modes is tracked at QwenLM/qwen-code#9511.

Install safety

A fresh Qwen 0.21.10 install into an isolated temporary home:

• installed and listed Cognee v1.4.4 with origin QwenCode
• loaded the context file and all four skills
• did not create ~/.qwen/settings.json
• did not modify the live Qwen extension, live Qwen settings, or Codex config (before/after hashes and metadata were identical)

Verification

• cd integrations/tests && uv run pytest tests/ -q — 1446 passed, 127 skipped, 53 deselected
• ruff format --check integrations/ — 308 files already formatted
• ruff check integrations/ — all checks passed
• Qwen JSON manifests/hooks validated with jq empty
• git diff --check origin/main...HEAD — clean
• independent full review plus scoped fix re-reviews — no remaining Critical or Important integration findings
• live credit-bearing suite was not run; its Qwen envelope behavior is covered hermetically

Related: #350
Qwen host prerequisite: QwenLM/qwen-code#9511

Closes #351

[netbrah]

@siillee :)