feat(openai): reconcile gpt-5.6 defs with upstream (cacheWrites + tiers)

[zephg]

What

Reconciles the fork's gpt-5.6 (Sol/Terra/Luna) model defs with upstream, which independently shipped the same three models after the fork did (so the entries now collide and differ).

• Adopt from upstream: cacheWritesPrice (Sol 6.25, Terra 3.125, Luna 1.25) and flex/priority service tiers. Without these the fork under-counted cost (no 1.25× cache-write price, no flex/priority tiers).
• Keep from the fork: Luna contextWindow: 1_050_000 + longContextPricing, the reasoning-effort defaults (Sol high / Terra medium / Luna low), and the descriptions.
• All tier contextWindows pinned to 1_050_000 (upstream tied them to its 400K Luna figure — still unverified against OpenAI docs).

Why

The fork shipped gpt-5.6 in #3 (ded80951d); upstream then shipped its own (a441e02bd). This cherry-pick gets both the cost fields and the fork's intended behavior right, and shrinks the guaranteed conflict on the next upstream sync.

Files

• packages/types/src/providers/openai.ts — add cacheWritesPrice + tiers to the three entries.
• src/api/providers/__tests__/openai-native.spec.ts — assert Sol cacheWritesPrice 6.25 + a priority tier.
• FORK.md — record the collision, the index.ts default-model divergence (openAiNativeDefaultModelId vs upstream's hardcoded "gpt-4o"), next-sync guidance, and GPT-5.6 provider coverage.

Verification

• pnpm --filter @roo-code/types check-types ✅
• openai-native.spec ✅ 53/53 (incl. new assertions); fetchers/openrouter.spec ✅
• reasoning.spec + model-params.spec — 22 failures = unchanged fork baseline (stale upstream budget/binary fixtures), no new failures
• pre-commit prettier + turbo lint ✅

Provider coverage (follow-on request — verified 2026-07-14)

• Bedrock: GPT-5.6 GA on Bedrock 2026-07-13, but Responses-API-only via bedrock-mantle; the fork's Bedrock provider is Converse-only, so a registry entry would 400. Deferred — needs a separate mantle-Responses transport.
• Vertex: GPT-5.6 is not offered by Vertex Model Garden (proprietary GPT ships to Azure/Bedrock). Claude on Vertex (Opus 4.8, Fable 5, Sonnet 5) is already correctly wired — no change.

Follow-up

Merge local/gpt-5.6-reconcile → local/daily-driver once reviewed (per the PR #3 workflow).

🤖 Generated with Claude Code