fix: allow max reasoning for gpt 5.6

[dirtydishes]

Thinking Path

Hermes already accepts max, and GPT-5.6 supports it natively. The WebUI’s OpenAI-family ceiling currently strips max from every GPT-5 model, so Sol, Terra, and Luna stop at xhigh even though the models can go higher.

There are broader PRs open for GPT-5.6 and ultra, but this one is intentionally boring: no new vocabulary, no catalog changes, no transport changes, and no guessing based on a loose gpt-5.6-* prefix. It only lifts the existing ceiling for the six model IDs Hermes currently knows support max.

What Changed

• Added an exact allowlist for GPT-5.6 Sol, Terra, and Luna, including each -pro variant.
• Preserved max for those six IDs on OpenAI-family lanes.
• Kept the existing ceilings for:
  • older GPT-5 models
  • o-series models
  • unknown GPT-5.6 variants
  • Gemini
  • pre-adaptive Claude
• Added focused coverage for model/provider hints, configured effort lists, status metadata, coercion, and gateway streaming.

Why It Matters

The WebUI should not quietly cap a model below a reasoning level the runtime supports. At the same time, capability policy should fail closed instead of handing max to every future slug that happens to begin with gpt-5.6-.

Contract Routing

Task type: reasoning capability-policy correction.

Touched areas:

• Server-side model capability filtering
• Reasoning status metadata
• Runtime effort coercion, covered through existing call paths

Relevant public docs:

• AGENTS.md
• CONTRIBUTING.md
• docs/GUIDELINES.md
• docs/CONTRACTS.md
• docs/UIUX-GUIDE.md
• DESIGN.md

Scope boundaries:

• No ultra level added.
• No model catalog changed.
• No frontend layout changed.
• No provider outside the existing OpenAI-family policy changed.

Verification

• Focused GPT-5.6 policy suite: 58 passed
• Reasoning and gateway slice: 424 passed
• Python compilation: passed
• Diff-scoped Ruff: no findings on added or modified lines
• git diff --check: passed
• Fresh independent review of exact commit/diff: no actionable findings
• Fresh /thermo-nuclear-code-quality-review: APPROVE, no CORE, SILENT, or NIT findings
• Full repository suite: 13,324 passed, 104 skipped, 1 xfailed, 2 xpassed; 115 failed in unrelated auth/workspace/TLS clusters
• Representative auth, workspace, upload, and TLS failures were rerun at pristine base dd554c17; auth and TLS failures reproduced unchanged, while the sampled workspace/upload tests passed independently (the larger full-suite cluster is state/order-sensitive)

Exact reviewed revision:

• Commit: f8f35b0ab4208f6a3b4c15582e497f4b342737c9
• Base: dd554c1788948ca5892f549575daf741aa6a8850
• Binary diff SHA-256: 6541d2866d513bef47ff449384d7ec2da2c5d824666df8a72f26a26b5550205d

Risks / Follow-ups

• The exact allowlist is deliberately fail-closed. New GPT-5.6 variants will need an explicit capability update rather than inheriting max by name.
• This overlaps conceptually with Add GPT-5.6 models and reasoning modes #5879 and fix(reasoning): expose max and ultra for GPT-5.6 #6018 but avoids their broader catalog/ultra scope.

Release note: GPT-5.6 Sol, Terra, Luna, and Pro variants can now use Max reasoning in the WebUI.

Model Used

• OpenAI Codex gpt-5.6-sol for implementation orchestration, with separate fresh independent and thermo-nuclear review agents.