Studio: share GGUF recommendation policy across pickers

[ousamabenyounes]

Summary

• share one GGUF recommendation policy across the Hub download card and chat model picker
• prefer the default variant when it fits, otherwise the largest fitting variant, then the smallest all-OOM fallback
• preserve each surface's existing fit thresholds by injecting its own fit predicate
• show the Hub recommendation on the collapsed trigger and beside the corresponding row when expanded, while suppressing the badge for downloaded or partial variants

Closes #7646

Test verification (RED → GREEN)

RED on the previous PR implementation: a downloaded fitting Q4_K_M was skipped and the remaining OOM Q8_0 was incorrectly recommended:

not ok - keeps a downloaded fitting variant as the recommendation instead of promoting an OOM download
expected: Q4_K_M (downloaded, fitting)
actual:   Q8_0 (OOM)

GREEN after the patch:

• focused recommendation suite: 10/10
• helper coverage: 100% lines, branches, and functions
• full frontend suite: 194/194
• TypeScript typecheck and production build pass
• Chromium/Playwright: verified the Hub card on the collapsed trigger and corresponding expanded row; verified the chat picker recommends a fitting variant rather than an OOM variant

The chat-picker integration test parses the executable TSX AST. An adversarial run with the expected expression present only in a comment fails, while the restored executable predicate passes.

AI assistance disclosure

This contribution was assisted by OpenAI Codex. I reviewed and understood the final patch and verified it with automated tests and Chromium.

[ousamabenyounes]

Updated the patch to share only the recommendation policy while preserving each picker’s existing fit thresholds. Added the downloaded-fit versus OOM regression case, fully structural chat-picker integration coverage, 100% helper branch coverage, and Chromium checks for both surfaces. Local validation: 10/10 focused tests, 194/194 full frontend tests, typecheck, and production build.