fix(codex): inject token saver prompts as instructions

[Jordannst]

Summary

Fixes Token Saver/Caveman prompt injection for Codex/OpenAI Responses requests.

When Token Saver injected a system prompt into a Responses-style request with no existing instructions field, it fell back to inserting a chat-style message into input[]:

{ role: "system", content: "..." }

Codex GPT-5.6 rejects that shape because Responses input[] entries must use Responses item/content schemas, so requests failed with:

Unknown parameter: 'input[0].content'

Changes

• Pass the target request format into injectMessagesSystem.
• For openai-responses, openai-response, and codex formats, append Token Saver/Caveman prompt text to top-level instructions instead of mutating input[].
• Preserve existing chat-completions behavior: non-Responses OpenAI-shaped requests still receive a system message in messages[].
• Add regression coverage for:
  • Responses requests with no existing instructions field.
  • Responses requests with existing instructions.
  • Chat completions messages[] injection behavior.

Verification

npx --yes vitest@4.1.10 run --environment node tests/unit/system-inject.test.js tests/unit/caveman-prompts.test.js

Result:

Test Files  2 passed (2)
Tests       12 passed (12)

Closes #2497

[t101804]

this thing gotta be merged asap

[dancju]

maintainers, merge this please