fix(translator): pass output_config.effort=max through verbatim

[ikelvingo]

Summary

Fixes a contract mismatch where output_config.effort="max" from an Anthropic-format request was incorrectly rewritten to xhigh when translating to an OpenAI-compatible upstream. This caused upstreams that accept max literally but reject xhigh — such as Ollama Cloud — to return HTTP 400 invalid reasoning value: 'xhigh'.

Root Cause

open-sse/translator/request/claude-to-openai.ts::normalizeOpenAIReasoningEffort unconditionally rewrote max → xhigh during step 1 (translation). However, provider-aware allowlist decisions belong in sanitizeReasoningEffortForProvider (the executor layer), which holds the full provider capability table — including providers that explicitly opt in to supportsMaxEffortForProvider(): ollama-cloud, opencode-go deepseek, moonshot k3, and native claude.

Because the translator runs before the sanitizer, the carrier was already xhigh by the time the sanitizer saw it, so the effortStr === "max" branch could never fire — the max information was lost before reaching the sanitizer.

Fix

Removed the if (normalized === "max") return "xhigh" line in claude-to-openai.ts. The translator now only performs form conversion (lowercase + non-empty filtering) and no longer makes allowlist decisions; max is written verbatim to result.reasoning_effort. Also updated the top comment in reasoningEffort.ts (English) to document this division of responsibility.

Impact

• Fix: Anthropic clients → OpenAI-shape upstreams that accept max literally (e.g. ollama-cloud) now pass max through unchanged instead of returning 400.
• No regression: the sanitizer's effortStr === "max" branch already covered the "client sends max directly in OpenAI-format" case; it now also covers the new "translator passes max through" path.

Verification

• tests/unit/translator-claude-to-openai.test.ts — assertion updated to expect max pass-through.
• tests/unit/base-executor-sanitize-effort.test.ts — added end-to-end test: translateRequest(CLAUDE→OPENAI) + sanitizeReasoningEffortForProvider("ollama-cloud", …) chain keeps output_config.effort="max" as max end-to-end.

Repro (before fix)

curl -X POST http://localhost:20128/v1/messages \
  -H "Content-Type: application/json" \
  -d '{
    "model": "ollama-cloud/gemma4:31b",
    "max_tokens": 4096,
    "messages": [{"role":"user","content":"hi"}],
    "output_config": {"effort": "max"}
  }'
# HTTP 400 {"error":{"message":"invalid reasoning value: 'xhigh'
#   (must be \"high\", \"medium\", \"low\", \"max\", or \"none\")"}}

After fix: HTTP 200, reasoning_effort: "max" forwarded to ollama-cloud.

[diegosouzapw]

Merged via local merge-train on 192.168.0.113 (32 cores) @ train tip e3a7696bb969536578f6a2defeada7d04fa9d1f9 — log /srv/omniroute-train/.claude/worktrees/merge-train-20260804-120425-suite.log.

Green: typecheck:core, check-complexity, check-cognitive-complexity, check-changelog-integrity, test:vitest (34 files / 291 tests).

Reds, each discriminated against the clean release tip 2e4268003 with zero PRs boarded (merge-gates §3):

• check-file-size — src/sse/handlers/chat.ts 1846>1845 and open-sse/executors/base.ts 1623>1578 reproduce with identical numbers on the bare tip. Baseline drift for the release captain; not introduced here.
• test:unit — 15 failures, all reproduced identically on the bare tip. Thirteen are the bare-model routing assertions (OpenAI remains the historical default, getModelInfoCore keeps unprefixed gpt-5.5 on the OpenAI fallback, …) left stale by fix(routing): bare model ids route to codex first; validate synced candidates #9275, which intentionally moved bare gpt-5.5/gpt-5.6-sol to codex without updating them; two are the trailing-period message change already covered by test(sse): expect the trailing period in the no-credentials message #9392.
• provider-limits-local-apikey-sync-spacing — the only failure not on the bare tip, and a load flake: the assertion requires a ~0ms gap and measured 25/49ms under a 32-core full-suite run. Passes 3/3 in isolation on this exact train tip.

No red attributable to this PR.