Tests for Responses path & reasoning fidelity

[HXYerror]

Part of #1. Final gate before merging the Responses-API epic.

Goal

Add a test suite covering the new Responses-API surface and the Anthropic↔Responses round-trip. Mirrors the existing tests under `tests/`:

• `tests/anthropic-request.test.ts`
• `tests/anthropic-response.test.ts`
• `tests/create-chat-completions.test.ts`

Tasks

• `tests/responses-request.test.ts`:
  • Validates request shape (`input`, `tools`, `reasoning`, `previous_response_id`)
  • Validates header construction (X-Initiator, Copilot-Vision-Request, copilot-integration-id)
  • Validates that `reasoning_effort` budget tiers map correctly when coming from Anthropic `thinking.budget_tokens`
• `tests/responses-response.test.ts`:
  • Output items array round-trip (reasoning + message + function_call interleaved)
  • `encrypted_content` is preserved (not stripped, not regenerated)
  • `status: null` is stripped
  • Streaming events: `response.output_text.delta`, `response.reasoning.delta`, `response.completed`
• `tests/anthropic-to-responses.test.ts`:
  • Anthropic `thinking` block on input → Responses `reasoning` item with preserved `encrypted_content` in `signature`
  • Responses `reasoning` item on output → Anthropic `thinking` block emission
  • Multi-turn round-trip preserves blob across turns
  • Tool-call interleaving correct (reasoning → tool_use → reasoning → text)
• `tests/model-routing.test.ts`:
  • `gpt-5.3-codex` on `/chat/completions` → 400 with helpful error
  • `gpt-4o` on `/responses` → translated to chat-completions or 400
  • `/v1/models` exposes `mode` field
• Streaming integration test: full SSE → Anthropic SSE round-trip with reasoning model

Acceptance criteria

• `bun test` passes with the new suite
• Coverage focused on the encrypted_content blob lifecycle (most fragile part)
• Regressions in existing chat-completions and `/v1/messages` tests caught

File pointers

• Existing tests under `tests/` (mirror their style)
• Bun test runner per `AGENTS.md`: `bun test tests/`

[HXYerror]

Implemented in commits 3ab9d01, b1133a3.

What was added:

• Full encrypted_content round-trip test (Responses → Anthropic → Responses)
  covering normal blob, no blob, empty-string blob, multiple simultaneous blobs
• Hard-fail assertions (throw instead of return) so critical signature check always runs
• Multiple reasoning items test: 2 blobs must preserve their distinct values
• model-routing: heuristic test now properly omits capabilities.type so heuristic fires
• o1-pro with capabilities.type:responses gets mode:'responses' from /v1/models

Production bugs fixed along the way:

• Empty-string encrypted_content was silently dropped by truthy guard —
  fixed to !== undefined in both translateReasoningItem and translateAssistantMessage
• thinkingBlock.signature comparison fixed to !== undefined (symmetrically)

220 tests passing across 13 files, lint clean.