fix(codex): prefer max_context_window over context_window as the usable input limit

[excessivechaos]

Problem

The live Codex OAuth catalog reports BOTH context_window (~272K) and max_context_window (~872K). The discovery parser's firstPositiveNumber chain reads context_window first, so every GPT-5.6 model resolves to a 272K input limit even though 272K is only the first pricing tier — the real usable window is max_context_window (872K).

Evidence that 272K is not a hard window: gpt-5.6-luna-xhigh served 380-390K input tokens with HTTP 200 (call_logs, 2026-08-20 02:45-02:48) while the 922K override was active — well past 272K with no context-overflow errors.

Consequence: the context-aware combo fallback treats codex as too-small for large conversations and reorders it behind 1M-context targets, defeating the declared priority order.

Fix

Reorder firstPositiveNumber so max_context_window wins over context_window when both are present; context_window remains the fallback for catalogs that omit the max.

Verification

• New regression test: live-shape payload (context_window: 272000 + max_context_window: 872000) resolves inputTokenLimit: 872000.
• End-to-end check with the real live payload: gpt-5.6-luna → 872000.
• Focused suite: 31 passed (provider-models-discovery-split).
• Changed-file ESLint + git diff --check clean.

Live payload keys verified directly against https://chatgpt.com/backend-api/codex/models with the account OAuth token (slug + context_window + max_context_window).

[excessivechaos]

Second commit (8520787) — the parser fix alone was insufficient: the conservative discovery merge (mergeCapacityLimitConservatively = Math.min(static, live)) still capped the live 872K at the static 272K from GPT_5_6_CODEX_CAPABILITIES. Verified by reproducing the merge in-process: live 872000 + static 272000 → merged 272000.\n\nThis commit raises the static Codex OAuth catalog to contextLength: 872000, maxInputTokens: 872000 (the usable max_context_window), matching the API-block precedent (GPT_5_6_API_CAPABILITIES uses 1.05M/922K). The live value now passes through the merge: live 872000 + static 872000 → merged 872000.\n\nAlso updated codex-gpt56-catalog.test.ts (272000 → 872000). Kiro tests unaffected (separate provider, still 272K).