Launch gate: hidden shapes and tolerance matrix

[metaforismo]

Goal

Create the hidden correctness suite and dtype-specific tolerance matrix for KernelForge.

Launch checklist

• Add hidden shapes for RMSNorm, RoPE, causal attention prefill, and KV decode.
• Include edge cases: near-zero RMSNorm rows, repeated keys, long-ish sequences, and uneven score distributions.
• Define dtype policy for float32, float16/bfloat16 where supported, and Python smoke floats.
• Define max absolute and relative tolerances per primitive/dtype.
• Add invalid-pattern checks for hardcoded public cases.
• Add hidden-case hash metadata without exposing values.
• Add verifier output fields for primitive-level pass/fail.

Acceptance criteria

• A candidate must pass hidden correctness before timing counts.
• Tolerance decisions are documented before official entries open.
• Hidden cases cannot be inferred from public smoke data.