Anthropic provider: structured-output schema rejected (minimum/maximum on integer/number) #76

Summary
In v2.1.4, the Anthropic provider fails on every LLM-mode scan: the structured-output
schema it sends includes minimum/maximum, which the Anthropic structured-output API
rejects.

Environment
SkillSpector 2.1.4 (from source, main) · Python 3.13 · SKILLSPECTOR_PROVIDER=anthropic ·
LLM mode (i.e. not --no-llm).

What happens
The semantic analyzers fail with two successive HTTP 400s from the Anthropic API — it
rejects minimum/maximum, once on an integer field and once on a number field.

Root cause
The output models declare start_line: int = Field(ge=1) and
confidence: float = Field(ge=0, le=1). Pydantic serializes these to JSON-Schema
minimum/maximum, and Anthropic's structured-output endpoint does not accept those
keywords on integer/number types.

Repro
Run skillspector scan <path> in LLM mode with SKILLSPECTOR_PROVIDER=anthropic and a
valid ANTHROPIC_API_KEY (no --no-llm). → two 400s, analyzers fail. Static --no-llm
mode is unaffected.

Suggested fix
Keep the Pydantic constraints (they're valid validation) and sanitize the schema sent to
Anthropic — strip minimum/maximum from the structured-output schema in the Anthropic
provider. (Confirmed locally that removing the constraints unblocks all three semantic
analyzers; schema-sanitization is the correct upstream form that preserves validation.)