fix(oracle): route maxTokens to maxCompletionTokens for GPT-5 family

[fede-kamel]

Summary

OCI's chat endpoint rejects maxTokens for the GPT-5 family with Use 'maxCompletionTokens' instead, so any GPT-5 request through the gateway currently fails with a 400 unless the caller already knows the OCI quirk and works around it.

This PR detects the GPT-5 family inline (/^openai\.gpt-5/i) and rewrites the parameter on the OCI chat request envelope. Everything else (gpt-4o, Llama, Cohere, Grok, Gemini) continues to use maxTokens unchanged.

Split out from #1537 per @narengogi's review.

What changed

src/providers/oracle/chatComplete.ts:

• Add a small inline usesMaxCompletionTokens(model) helper (no separate modelConfig.ts module — kept tight).
• In the chatRequest transform, after building the OCI envelope, swap maxTokens → maxCompletionTokens when the model is in the GPT-5 family.

src/providers/oracle/chatComplete.test.ts:

• openai.gpt-5 and openai.gpt-5-mini route to maxCompletionTokens.
• openai.gpt-4o, meta.llama-3.3-70b-instruct, cohere.command-r-plus-08-2024, xai.grok-3-mini, google.gemini-2.5-flash keep maxTokens.

Diff: +63 / −1, two files.

Verified

• npm run build ✓
• npm run format:check ✓
• Type-clean (tsc --noEmit no errors in src/providers/oracle/)
• Verified live against inference.generativeai.us-chicago-1.oci.oraclecloud.com/20231130/actions/chat for openai.gpt-5 (now succeeds; previously returned 400 "Use 'maxCompletionTokens' instead").

Related

• Split from fix(oracle): emit tool_calls in streaming responses #1537 per @narengogi's review
• Supersedes feat(oracle): Add rerank and guardrails support #1540 (the GPT-5 portion). This PR uses an inline regex (/^openai\.gpt-5/i) instead of feat(oracle): Add rerank and guardrails support #1540's separate modelConfig.ts module — same behaviour, smaller surface area.
• Sibling: fix(oracle): emit tool_calls in streaming responses #1537 (streaming tool-call fix), feat(oracle): add Cohere embeddings support #1634 (Cohere embeddings)

[fede-kamel]

@roh26it @narengogi @VisargD — companion to #1537. Rebased on main, mergeable, e2e-verified live against openai.gpt-5 on us-chicago-1 (HTTP 200; without this fix OCI returns 400 "Use 'maxCompletionTokens' instead"). Diff is +63 / −1, two files. See #1537 for the broader context. Happy to take review feedback.

[fede-kamel]

Friendly review ping @narengogi — this is the GPT-5 maxCompletionTokens split you asked for back when #1537 was trimmed. Description updated to reflect that it supersedes #1540's GPT-5 portion (this one uses an inline regex; the old one had a separate modelConfig.ts module). CI green, mergeable, ~63 lines + tests.