Add automatic Charm Hyper model sync

[JavaGT]

Summary

• add a registered Charm Hyper sync adapter
• derive reasoning effort levels, limits, modalities, pricing, and cache pricing from the live catalog
• preserve canonical base-model metadata and skip unknown models safely

Verification

• bun models:sync hyper --dry-run reports 20 unchanged
• bun validate
• git diff --check

[JavaGT]

Closing as duplicate of #3352, which already contains the Charm Hyper provider and sync adapter. Please review/fix the adapter issues identified by the automated review there, especially reasoning fallback when metadata is absent, discovery of newly listed models, provider auth env metadata, and cache pricing semantics.

[github-actions]

Action items

• [high] [possible mistake] packages/core/src/sync/providers/hyper.ts:71 - Check: Provider reasoning must reflect whether the model reasons on this provider; missing controls use reasoning_options = [], not reasoning = false. Why: reasoning is set true only when the catalog reasoning object is non-null. That forces reasoning = false on models whose base metadata (and this PR’s earlier hand-authored reasoning_options = []) mark them as reasoning models—e.g. kimi-k2.5, kimi-k2.7-code, minimax-m2.7, glm-5, gemma-4-26b-a4b-it, and the Qwen3.6/3.7 entries—so the catalog claims they do not reason. Action: Treat a missing/null reasoning object as “no verified control” (inherit base reasoning, or keep reasoning = true with reasoning_options = []). Only set reasoning = false when Hyper explicitly documents a non-reasoning deployment.
• [medium] [violation] packages/core/src/sync/providers/hyper.ts:72 - Check: Sync pricing should emit stable per-1M costs like other providers. Why: Raw API floats are written unchanged, producing values such as 0.20000010875000002 / 0.40000021750000003 / 0.039999586250000004 in providers/hyper/models/deepseek-v4-flash.toml (and similar noise on glm/kimi models). Action: Round prices before emit (same approach as OpenRouter/DeepInfra/etc.) and re-sync so TOMLs store clean decimals.
• [medium] [violation] packages/core/src/sync/providers/hyper.ts:113 - Check: resolveBaseModel must map Hyper IDs to existing models/ paths for creates, not only updates. Why: minimax-m2.7 resolves to minimax/minimax-m2.7, but metadata is minimax/MiniMax-M2.7. Without an existing TOML base_model, sync skips the model, so a clean or post-delete sync cannot recreate it. Action: Add a BASE_MODEL_ALIASES entry (or case-aware lookup) for minimax-m2.7 → minimax/MiniMax-M2.7, and verify cold sync still creates the file.
• [medium] [possible mistake] packages/core/src/sync/providers/hyper.ts:75 - Check: pricing.cache_hit → cache_read and pricing.cache_create → cache_write must match Hyper’s field meanings. Why: After sync, many models only get a ~50% of input price on cache_write with cache_read = 0 (e.g. gpt-oss-120b, kimi-k2.5, llama-3.3-70b-instruct, minimax-m2.7), while earlier hand-authored data put those amounts on cache_read. That pattern looks like cache-hit discounts, not typical cache-write premiums (compare Qwen3.6 entries where both sides look normal). Action: Confirm Hyper’s cache_hit/cache_create semantics against docs or a sample payload; fix the mapping (or post-process) so read/write prices are not swapped, then re-sync.