fix: resolve omitted Optional inputs to None at the HTTP edge

[markphelps]

Problem

A predictor input declared Optional[T] / T | None is emitted as an optional field in the OpenAPI schema (nullable: true, absent from required), but whether it can actually be omitted at request time depended on a Python-level detail the schema doesn't capture: whether the parameter had a usable __defaults__ entry.

Signature	Omitted at runtime
value: Optional[str] (bare, no Input)	❌ TypeError: missing 1 required positional argument
value: Optional[str] = Input(description="…")	✅ resolves to None
value: Optional[str] = Input(default=None)	✅ resolves to None

All three generate a correct "optional" schema. The bare form was a silent footgun: it built and served fine, advertised the field as optional, then failed at prediction time whenever a caller omitted it.

Fix

Validation against the OpenAPI schema already runs at coglet's HTTP edge (InputValidator), not in the worker. After validation passes, inject an explicit null for every property that is nullable: true, not in required, and has no default key, when it is absent from the request.

This rule mirrors the schema-generation discriminator in pkg/schema/openapi.go (buildInputSchema), keeps the OpenAPI schema as the source of truth for what is optional, and leaves the worker unchanged (it needs no schema).

The rule safely skips:

• Optional[T] = Input(default=None) — emits default: null, already handled by unwrap_field_info_defaults
• real defaults like Input(default=42) — emits default: 42
• required fields — the validator rejects omission first

Injection runs only after successful validation, so missing required fields still return 422. Training inputs (TrainingInput) are covered via the same path. Union optionals (A | B | None) are forced into required by the generator and are correctly skipped (they 422 on omission rather than resolving to None — pre-existing behavior, not a regression).

Tests

• Unit (input_validation.rs): inject for bare optional / optional list / optional enum; skip explicit-default and required fields; never override a present value.
• Unit (service.rs): predict injects after Ok, does not inject when a required field is missing, train path injects.
• Integration (optional_input_no_default.txtar): a bare value: Optional[str] omitted via both cog predict and POST /predictions {"input":{}} resolves to None.

[ask-bonk]

LGTM

github run