refactor(proxy): audit thinking-mode protocol and refactor test suite

[yxlao]

TL;DR

This PR audits the proxy against DeepSeek's official thinking-mode tool-call protocol and confirms that the multi-round reasoning_content chain is reconstructed and forwarded correctly across tool-call turns. Along the way it tightens a few protocol edges, fixes minor bugs, and consolidates the test suite.

Summary

Tightens the thinking-mode protocol behavior based on a code audit, and refactors the test suite around a single end-to-end protocol harness so each scenario lives in exactly one place.

Changes Made

Protocol fixes

• src/deepseek_cursor_proxy/transform.py:
  • Added strip_recovery_notice_for_upstream() so the proxy-generated recovery notice is no longer echoed back to DeepSeek. The notice still serves as a boundary marker in the with-prefix history (preserved for cache-scope alignment), but is stripped before the request leaves the proxy.
  • Removed LEGACY_RECOVERY_NOTICE_TEXT constant and its has_recovery_notice() branch — only the current recovery notice is recognized now.
  • Added WARNING logs for unsupported request fields (parallel_tool_calls, service_tier, etc.) that the proxy silently drops, and for non-DeepSeek model names that get rewritten to the configured upstream model.
  • Reordered rewrite_response_body(): prefix → record reasoning → fold reasoning, so cache scope signatures still match what Cursor will echo back.
• src/deepseek_cursor_proxy/streaming.py: Added fold_reasoning_into_content() for non-streaming responses, mirroring reasoning_content into a <details><summary>Thinking</summary>...</details> block (or <think> if collapsible_reasoning=False) — the streaming adapter already did this; non-streaming responses now match.
• src/deepseek_cursor_proxy/config.py + server.py: Removed the pass-through thinking mode. The proxy was the only thing that knew how to rebuild missing reasoning_content; pass-through bypassed it and produced 400s upstream.

Test suite refactor

• Consolidated overlapping tests into one canonical end-to-end file, tests/test_protocol.py, which uses a StrictFakeDeepSeek HTTP handler that 400s on missing reasoning_content (the same behavior the real DeepSeek API has). Test classes cover: canonical loop, strict-reject mode, thinking-disabled, recovery, streaming-then-non-streaming, streaming/non-streaming display, concurrent threads, and streaming cache timing.
• Removed tests/test_proxy_end_to_end.py (1414 lines) — its scenarios now live in test_protocol.py.
• Trimmed tests/test_transform.py from 1489 → 321 lines: kept only pure-helper unit tests (content extraction, request prep, recovery-notice stripping, response rewrite). Behavior that needs an upstream now lives in test_protocol.py.
• Expanded tests/test_server.py with HTTP-boundary tests: bearer token forwarding, oversized body rejection, healthz, streaming close-after-[DONE], normal vs verbose logging.
• Expanded tests/test_trace.py with three integration tests that exercise the trace writer through a running proxy (non-streaming replay, streaming chunks, recovery diagnostics).
• Added tests/test_streaming.py::FoldReasoningTests for the new non-streaming fold helper.

Final layout: 80 tests across 9 files (~3500 lines), down from 98 tests across 10 files (~5000 lines).

Repo cleanup

• Removed docs/audit_report.md and docs/thinking-tools.md — internal audit notes, not user-facing docs.
• Removed scripts/audit_deepseek_protocol.py — the mock DeepSeek server is now part of tests/test_protocol.py.
• Removed .claude/settings.local.json and added .claude/ to .gitignore.
• Dropped the now-obsolete scripts/audit_deepseek_protocol.py ruff per-file ignore from pyproject.toml.

Breaking Changes

• --thinking pass-through is no longer supported. Configs that specified pass-through will fall back to the default (enabled). The mode bypassed the proxy's reasoning-cache patching, which is the proxy's reason for existing.
• The legacy recovery-notice prefix ("Note: recovered this DeepSeek chat...") is no longer recognized as a boundary marker. Live conversations that still carry the old prefix in their history will have it treated as ordinary assistant content. New recoveries always emit the current [deepseek-cursor-proxy] Refreshed reasoning_content history. prefix.

Test plan

• uv run python -m unittest discover -s tests — 80 tests pass (1 skipped: live DeepSeek API test, requires real API key)
• Manual: run the proxy against Cursor with a multi-turn tool-call conversation, verify no 400s
• Manual: kill and restart the proxy mid-conversation, verify the recovery notice appears once and subsequent turns succeed