[autoantohaki]

Requested by mau

Before: a gRPC request that set only some Sampling fields was rejected. With proto3 scalars, an unset top_p arrives as 0 and an unset max_tokens as 0, so {"sampling":{"maxTokens":32}} failed with INVALID_ARGUMENT: top_p must be greater than 0, and {"sampling":{"seed":7}} failed on max_tokens. Only "no sampling at all" or "every field set" worked.

After: each unset field falls back to the same default HTTP uses (max_tokens 256, temperature 0.8, top_p 0.95), so callers can set just what they need. An explicit temperature: 0 still means greedy decoding.

How: max_tokens, temperature, and top_p are now proto3 optional (wire-compatible: existing clients that send values are unaffected), and sampling() in grpc.rs defaults each field independently. A unit test covers partial, greedy, and absent sampling; the gRPC API doc notes the defaults.

CI needs #1 merged first (it restores the privacy-scan and supply-chain checks on main).