Distinguish gpt-5.3-codex-spark from base GPT-5.3 Codex label

[ozymandiashh]

Summary

Closes #461 by giving gpt-5.3-codex-spark its own display label instead of collapsing it onto the retired base GPT-5.3 Codex in today/models/status output.

Root cause

getShortModelName (src/models.ts) and the Codex provider's modelDisplayName (src/providers/codex.ts) resolve display names with a longest-key-first startsWith(key + "-") match. That match is intended for reasoning-effort suffixes (-high/-low), which should fold onto the base label. But -spark is a distinct model variant, not a reasoning suffix, and there was no explicit entry for it, so it was swept onto gpt-5.3-codex and rendered as GPT-5.3 Codex. On usage/billing dashboards this makes a distinct, active model look like the retired base and muddies spend attribution.

As confirmed on the issue, pricing was never wrong: LiteLLM has no gpt-5.3-codex-spark entry, so getModelCosts falls back to the base gpt-5.3-codex rate. This is a display-label-only defect.

What changed

• Add an explicit display entry gpt-5.3-codex-spark -> GPT-5.3 Codex Spark to SHORT_NAMES in src/models.ts.
• Add the matching entry to modelDisplayNames in src/providers/codex.ts.

Both tables are evaluated longest-key-first (SORTED_SHORT_NAMES in src/models.ts, modelDisplayEntries in src/providers/codex.ts), so the longer, more specific key wins for the exact variant while -high/-low continue to fall through to the base. The change is purely additive: no other model label or code path is touched.

Validation

Concrete before/after via the real getShortModelName resolver:

// origin/main (5fd1160)
gpt-5.3-codex-spark => GPT-5.3 Codex
gpt-5.3-codex-high  => GPT-5.3 Codex
gpt-5.3-codex-low   => GPT-5.3 Codex
gpt-5.3-codex       => GPT-5.3 Codex

// this branch
gpt-5.3-codex-spark => GPT-5.3 Codex Spark
gpt-5.3-codex-high  => GPT-5.3 Codex
gpt-5.3-codex-low   => GPT-5.3 Codex
gpt-5.3-codex       => GPT-5.3 Codex

Only gpt-5.3-codex-spark changes; the reasoning suffixes and the base are untouched.

Checks run (en_US locale):

npm test -- --run tests/models.test.ts tests/providers/codex.test.ts   - 128 tests passed
npm test -- --run                                                      - 1276 tests passed
npx semgrep --config .semgrep/rules/no-bracket-assign-hot-paths.yml src/  - 0 findings (37 files)
npm run build (bundle-litellm + tsup)                                  - build success
Gemini 3.1 Pro (High) review - PASS (sort is order-independent; no collateral relabel:
  `-sparkle` falls through, `-spark-high` maps to Spark; pricing unaffected; no orphaned display paths)

New regression coverage:

• tests/models.test.ts: gpt-5.3-codex-spark -> distinct label; -high/-low -> base.
• tests/providers/codex.test.ts: same two assertions through the Codex provider's modelDisplayName.

Notes

• Pricing/cost is unchanged by design (no LiteLLM key for the variant; base-rate fallback preserved).
• npm run build regenerated src/data/litellm-snapshot.json and src/data/pricing-fallback.json; both were restored because they are unrelated to this fix. The diff is the two source entries plus tests.