feat: gpt-5.5 + claude-opus-4-7-max + gpt-5.3-codex tier ladder (v2.0.50)

GetCascadeModelConfigs response on 2026-04-30 surfaced new SKUs that
were only reachable through mergeCloudModels' dynamic ingestion path:

  - gpt-5-5-{none,low,medium,high,xhigh} + matching -priority lane
  - claude-opus-4-7-max (above xhigh in the effort ladder)
  - gpt-5-3-codex-{low,high,xhigh} + matching -priority

The implicit path lost dotted aliases (gpt-5.5-medium), kept only the
ugly cloud-format names (gpt-5-5-medium), and used cloud's default
credit values rather than curated ones. issue #109 reported this.

Mirror the gpt-5.2/5.4 pattern: explicit MODELS entries with curated
credits + bidirectional alias coverage (cloud format <-> dotted format).
Default bare gpt-5.5 to medium tier per the existing 5.2/5.4 convention.

Also clean up swe-1-6 / minimax-m2-5 / kimi-k2-5 cloud-format aliases
so mergeCloudModels stops creating duplicate entries when the cloud
sends both formats in the same response.

Tests: 497 -> 500 (+3). Regressions pin the full 5.5 ladder, opus-max
alias, and 5.3-codex tier coverage.