fix(pricing): restore the DeepSeek v4 and gpt-5.6-codex repricing invariants (#1134)

[ozymandiashh]

Closes #1134.

What this does

The 0.9.21 snapshot refresh caught the 2026-08-24 repricings mid-transition and relaxed two tests/models.test.ts invariants ("cannot be asserted until upstream syncs"). This PR restores both, now that upstream has synced, and fixes a bundler ordering defect the refresh exposed.

DeepSeek v4 — prefixed-equals-bare restored

LiteLLM now carries the official peak rates for both spellings (pro $1.32/$3.96 per million, flash $0.30/$1.20, cache-read $0.044/$0.006 — verified against api-docs.deepseek.com/quick_start/pricing, off-peak is half of peak). The hand-pinned MANUAL_ENTRIES for deepseek-v4-flash/deepseek-v4-pro (added while LiteLLM PR #27056 was still open) are removed: they shadowed upstream with the pre-sync promotional rates. deepseek/deepseek-v4-pro now equals deepseek-v4-pro, and the "official pricing" tests pin the current peak rates.

gpt-5.6-codex / -max — codex-equals-base restored

The codex SKUs are still absent from LiteLLM as of the 2026-09-29 refresh, so their MANUAL_ENTRIES remain — but now they mirror the repriced gpt-5.6 base row verbatim, tier block included ($4/$20 per million, 1.25x cache-write, >272k tier at 2x), per the same-generation pattern every codex id LiteLLM carries follows. The toEqual(snapshot['gpt-5.6']) assertion holds again, and if the base row moves, the mirror fails the test until refreshed. The long-context tier test now expects the codex SKUs to carry the base tier (previously undefined, because the flat manual tuple had no tier slot).

Bundler fix: reseller rows no longer shadow official rows

Pass 2 of scripts/bundle-litellm.mjs lets a prefixed row claim the slot its vendor-stripped key would answer. LiteLLM sorts reseller namespaces before vendor rows (openrouter/deepseek/deepseek-v4-pro at index 2167, deepseek/deepseek-v4-pro at 3375), so the openrouter resale row claimed the deepseek/deepseek-v4-pro slot first and the official row could never take it back (fillsOnly vetoes re-pricing) — the winner was decided purely by upstream JSON key order, and the namespaced id priced at the resale rate, ~40% under official peak. A prefixed row may now claim only a stripped slot that is not itself an upstream entry name; multi-segment ids that exist only as reseller rows (e.g. accounts/fireworks/models/...) keep resolving exactly as before.

Refreshed snapshot

Regenerated with npm run bundle-litellm: besides the deepseek/codex corrections, 24 rows moved by ordinary upstream drift (claude-sonnet-5-5 family added across providers, deepseek-v4.1-flash repriced, zai/glm-4.6 cache-write slot filled, two dropped mistral embed rows carried forward). Token-priced calls re-price on read, so no re-parse is needed.

Testing

• tests/models.test.ts + tests/pricing-fallback-data.test.ts: 212 passed
• tests/cli-deepseek-v4-pricing.test.ts: updated to the official peak rates (same observed token counts, recomputed dollars)
• Full vitest run green apart from two pre-existing environment flakes (missing react-dom/server in the app workspace, one timing-sensitive electron retire test that passes standalone)