feat(models): add GPT-6.1 Sol catalogs, pricing, and Copilot routing

[HouMinXi]

Summary

GPT-6.1 Sol is in the live Codex catalog, but the built-in OpenAI and Codex tables on release/v3.8.52 do not list it. GitHub Copilot's static catalog stops at gpt-5.6. The Copilot executor only sends model names containing codex to /responses, so gpt-6-astra and gpt-6.1-sol hit /chat/completions and fail.

This adds the base ids to OpenAI, Codex, GitHub Copilot, and GHE Copilot, and sends any gpt-6* Copilot model to /responses.

Long-context billing follows the published Sol page: once input is above 272,000 tokens, the whole request is charged at 2x input, 2x cache, and 1.5x output. The bare gpt-6-sol id is unchanged. The reasoning premium uses the scaled output price and is clamped at zero, because reasoning tokens are already counted in the output total.

Public API limits stay 1,050,000 context / 128,000 output. Codex's live maximum context (872,000) is kept separate. max is preserved through translation.

The fallback Codex client pin moves from 0.156.1 to 0.159.2, in both the header default and the image CLI (@openai/codex). A caller-supplied version still wins. Compared with 0.156.1, 0.159.2 does not add request headers: originator stays codex_cli_rs, the User-Agent format string is unchanged, and the session headers remain session-id and thread-id. The public API default effort stays medium.

Review

Forge LOCAL on this change did not reach three clean rounds. The confirmed findings on the last three runs were excerpt mismatches (<receipt-evidence>), not product defects. Product notes that were real (reasoning premium ignoring the scaled output price, then going negative) are fixed in this commit. Repeated notes about the cache rate 0.2, the gpt-6-sol bare id, and \d+ vs gpt-6.10 were checked against the code and rejected. The client-pin bump landed after that review; it is a version string plus the tests and snapshot that lock to it.

Test plan

• tests/unit/codex-gpt6-sol-luna.test.ts — catalog ids, effort suffix, long-context boundary at 271,999 / 272,000 / 272,001
• tests/unit/reasoning-cost-double-billing.test.ts
• tests/unit/8951-github-gpt56-responses.test.ts — gpt-6-astra and gpt-6.1-sol resolve to /responses
• tests/unit/claude-codex-identity-version-sync.test.ts, client-identity-profiles, codex-client-headers, executor-codex — pin matches the Dockerfile

Related to #15166

[JxnLexn]

Related: my PR #15132 makes the live Codex account catalog authoritative for the dashboard and /v1/models. After a successful sync, a new base model such as gpt-6.1-sol can appear without a static registry addition. Static entries are retained for metadata enrichment and an explicit offline/bootstrap fallback, but static-only IDs and reasoning-suffix aliases are not appended to the synced list.

So there is some overlap with the model-picker part of this PR, but this is not a duplicate overall: the pricing, Fast-tier, reasoning handling, and GitHub Copilot changes here are outside #15132's scope. It would be good to keep those additions compatible with the live catalog, particularly so the new effort aliases do not reappear as separate selectable models after synchronization.

The companion PR #15133 separately selects accounts using their own synced model inventories, so availability on one account does not imply availability on every account. Both PRs target release/v3.8.52.

[HouMinXi]

Superseded by #15192 (a3401e9), which landed GPT-6.1 Sol on release/v3.8.52. The codex registry on the tip already lists gpt-6.1-sol and its tier variants.