fix(sse): re-run strict system hoist after format translation

[Kizuno18]

Setting OMNIROUTE_STRICT_SYSTEM_PROVIDERS wasn't enough to get Claude Code talking to a strict upstream, and the reason turned out to be ordering rather than configuration — the hoist runs before the target translator, and claudeToOpenAI puts a system message back in front of the one it just hoisted.

My setup: a vLLM endpoint serving Qwen3.8-27B, registered as a vllm connection, with OMNIROUTE_STRICT_SYSTEM_PROVIDERS=vllm. Qwen3's chat template rejects anything but a single system message at index 0 — same constraint xiaomi-mimo already has in BUILTIN_PROVIDERS_SYSTEM_MUST_BE_FIRST:

{"error":{"message":"System message must be at the beginning.","type":"BadRequestError","code":400}}

Claude Code sends a top-level system field and a role: "system" message inside messages (its deferred-tools/agents block, ~6.5KB, present on every request — it isn't a hook or a plugin, so there's no way to turn it off client-side). That's the shape that breaks:

• hoistLeadingSystemMessage at open-sse/translator/index.ts:321 runs on the source array and correctly folds the offender onto index 0
• then claudeToOpenAI (open-sse/translator/request/claude-to-openai.ts:161) pushes body.system as a fresh leading system message and appends the converted messages after it
• the hoisted system is back at index 1, and the upstream 400s

A Responses-source request never gets that far — result.messages doesn't exist yet at line 321 (only input does), so the call is a plain no-op and Codex CLI hits the same wall through developer → system normalization.

The fix re-runs the same helper on the final outbound array, at the single return of translateRequest, which is the only shape the upstream actually sees. It's the existing helper with no behaviour change: still merge-never-drop, still returns the same array reference for non-strict providers and already-compliant requests, so prompt-cache prefixes are unaffected.

Confirmed against the live endpoint — the array claudeToOpenAI emits today is rejected, and the array the helper produces from it is accepted with both texts preserved:

messages sent upstream	result
[system, user, system, user] (current)	400 System message must be at the beginning.
[system, user, user] (after the fix)	200, system content "You are a coding assistant.\ndeferred tools list"

Added a case to tests/unit/probe-7293-strict-system-hoist.test.ts covering the Claude → OpenAI path with both a top-level system and a mid-array one, asserting a single system at index 0 with both texts merged. Stashing the index.ts change and re-running gives 5 pass / 1 fail with only the new case failing, so it does pin the regression rather than pass vacuously. memory-system-first-6135, claude-system-role-cache-boundary and memory-cache-safe-injection are green too (33 assertions), and ESLint is clean on both files.

Worth a look at whether the pre-translation call at line 321 is still needed once this one exists — I left it alone to keep the diff small, but the later call may well subsume it.