Add gpt-5.5 model overlay (OpenAI, 2026-04-23 release)

[amargandhi]

Summary

OpenAI shipped GPT-5.5 on 2026-04-23. This PR adds the matching model overlay so gstack users on Codex CLI / Cursor / OpenCode running 5.5 get model-specific guidance baked into generated SKILL.md files.

Naming trap up front: the model ID is gpt-5.5, NOT gpt-5.5-codex. OpenAI collapsed the naming — one model sits at the top of both ChatGPT and Codex CLI. scripts/models.ts defensively aliases gpt-5.5-codex* → gpt-5.5 because users will type the suffix from muscle memory.

gpt-5.3-codex and gpt-5.3-codex-spark remain first-class — not superseded. No Cerebras variant shipped for 5.5.

What this adds

1. model-overlays/gpt-5.5.md (68 lines, inherits from gpt):

   • reasoning_effort ladder: low | medium | high | xhigh + when to use each
   • Migration gotcha: Codex 0.124.0+ resets reasoning_effort on model swap; users muscle-memorying codex -m gpt-5.3-codex -c 'model_reasoning_effort=\"xhigh\"' need to re-pass -c explicitly when swapping to 5.5
   • 1M API context / 400K Codex CLI throughput cap
   • Honest cross-provider read: where Opus 4.7 still edges 5.5 (SWE-bench Pro 64.3 vs 58.6, HLE no-tools 46.9 vs 41.4, MCP-Atlas 79.1 vs 75.3)
   • Skill-by-skill recommendations: when 5.5 wins (long-horizon agentic, 1M context retrieval), when to stay on 5.4 / 5.3-codex-spark (sub-second iteration, mechanical tasks)
   • Mentions auto-review hook stable in Codex 0.124.0 (pairs nicely with /review)
   • Hallucination caveat: AA-Omniscience reports 86%; verify aggressively for citation-heavy work

2. scripts/models.ts: adds gpt-5.5 to ALL_MODEL_NAMES and two heuristics (gpt-5.5-codex* → gpt-5.5, gpt-5.5* → gpt-5.5). Order matters: 5.5 matchers come first because the codex-suffix alias would otherwise fall through to the generic gpt matcher.

Test plan

• bun test test/skill-validation.test.ts test/gen-skill-docs.test.ts — 689/689 pass
• bun run gen:skill-docs --host all — clean regen, no warnings
• Default --host all generation unchanged (no model flag = no overlay baked, no per-skill drift)
• Built against upstream/main at v1.15.0.0 — minimal surface area, no preamble or skill template touched

Notes for review

• This PR is intentionally small. The other contributions on my fork (/challenge, /glossary, canonical engineering vocabulary) are larger and I'm happy to PR them separately if there's interest — wanted to start with the lowest-risk one.
• Some claims in the overlay are marked UNVERIFIED (e.g. default reasoning_effort value at launch, gpt-5.5-pro general API availability) — happy to update once OpenAI publishes definitive docs.