fix: retry transient HTTP 400 errors from upstream providers

[BlueBoobyAI]

Problem

Claude Code sessions crash multiple times per day when proxied through free-claude-code to DeepSeek V4 Flash:

Invalid request sent to provider. Request ID: req_d10fcb8d9fb7

The user sees a fatal error in their terminal:

Provider request failed unexpectedly.
Request ID: req_7317b1104a14

Root cause: DeepSeek occasionally returns HTTP 400 on transient internal hiccups (not real request bugs). The proxy's retryable_upstream_status() function at providers/rate_limit.py:28 only treats HTTP 429 and 5xx as retryable — 400 passes through immediately, raising InvalidRequestError → session crash → user restarts Claude Code.

The /messages POST is idempotent and 400s aren't billed, so retrying is safe.

Fix

providers/rate_limit.py — add if status == 400: return 400 to both branches of retryable_upstream_status():

# Before: 400 falls through to return None → immediate crash
if isinstance(exc, httpx.HTTPStatusError):
    status = exc.response.status_code
    if _upstream_http_retryable(status):  # only 429 + 5xx
        return status
    return None  # ← 400 dies here

# After: 400 enters the existing exponential-backoff retry loop
if isinstance(exc, httpx.HTTPStatusError):
    status = exc.response.status_code
    if _upstream_http_retryable(status):
        return status
    if status == 400:
        return 400  # ← transient 400 → retry
    return None

Important safety detail: HTTP 400 retries do NOT call set_blocked() on the shared GlobalRateLimiter. Unlike 429/5xx (which signal upstream congestion worth a global pause), a transient 400 is a per-request hiccup. A genuine bad request (wrong model name) retries with individual backoff but does not stall concurrent requests.

Additional fixes per review:

• retryable_upstream_status docstring updated to document 400 behavior
• Log label changed from "Upstream server error (400)" to "Transient bad request (400)" — 400 is a client error, not server error
• _upstream_http_retryable docstring notes 400 is intentionally excluded (separate branch to skip set_blocked)

Safety Evidence

All 1440 existing tests pass. New tests added:

Unit tests (test_provider_rate_limit.py)

• test_execute_with_retry_400_retried_then_exhausts — httpx HTTPStatusError with 400: asserts 3 calls (1 initial + 2 retries)
• test_execute_with_retry_400_then_200_recovers — transient 400 then 200: asserts call_count == 2 and result == "ok"
• test_execute_with_retry_openai_400_retried_then_exhausts — openai.BadRequestError with 400: asserts 3 calls

Integration test (test_anthropic_messages_429_retry.py)

• test_transient_400_is_retried_then_exhausts — real execute_with_retry, 5 send calls, SSE error envelope with "Invalid request sent to provider."