feat(providers): modernize Naver CLOVA Studio chat and embeddings

[backryun]

Summary

Modernize Naver CLOVA Studio as one provider change:

• migrate chat from the OpenAI compatibility shim to native Chat Completions v3
• add native CLOVA Embedding v2 behind the OpenAI-compatible /v1/embeddings surface
• keep the whole change in one commit and one PR

Chat Completions v3

clova-studio now targets /v3/chat-completions/{modelName} with a dedicated
clova wire format, OpenAI-to-CLOVA request translation, CLOVA-to-OpenAI response
translation, and a small specialized executor that owns the model-in-path URL rule.

Supported and live-verified models:

Model	Text	Vision	Reasoning	Tools	Structured output
HCX-005	yes	URL + base64	no	yes	no
HCX-007	yes	no	yes	yes	yes
HCX-DASH-002	yes	no	no	yes	no

The translator encodes the upstream constraints discovered during live verification:

• HCX-007 uses maxCompletionTokens and is the only model receiving thinking
• function calling on HCX-007 forces thinking.effort: "none"
• function-calling output caps below 1024 are raised to the upstream minimum
• only toolChoice: "auto" and "none" are forwarded
• public image URLs and full base64 data URIs are supported on HCX-005
• terminal result events do not repeat text/tool calls already emitted by token events
• upstream status envelopes are converted into sanitized provider failures

Embedding v2

The embedding registry now exposes:

clova-studio/clova-embedding-v2

The native endpoint accepts one { "text": ... } input per request and returns a
CLOVA envelope rather than OpenAI's data[] shape. OmniRoute therefore:

• validates non-empty text-only input
• fans an OpenAI batch out sequentially, preserving input order and response indexes
• normalizes every native response into OpenAI object: "list" output
• aggregates prompt/total token usage across the batch
• counts the real number of upstream requests for quota accounting
• rejects unsupported base64 encoding, token-array input, and non-1024 dimensions
• converts HTTP-200 CLOVA error or malformed envelopes into provider failures

Quality and regression work

• moved the model-path rule out of the already-large default executor into
  ClovaStudioExecutor
• split the embedding handler into bounded lifecycle helpers without changing the
  existing provider contracts
• updated the provider translate-path and specialized-executor golden snapshots
• registered the new translator pair and specialized executor with repository gates

Verification

• CLOVA, embedding, executor, and golden regression set: 139/139 passed
• Vitest: 50 files, 463 tests passed
• npm run typecheck:core: passed
• open-sse typecheck: 0 errors
• API frozen-baseline gate: passed with no new errors
• complexity new-code gate: 0 cyclomatic / 0 cognitive violations
• provider consistency: passed
• known-symbols: passed
• scoped ESLint and formatting: passed

Scope note

CLOVA's reranker remains intentionally excluded. It behaves as a RAG answer generator
without relevance scores and does not satisfy OmniRoute's /v1/rerank contract.