[SDK gap] No shared canonical PCM / conformance helper — every language reimplements the same hash

[balloob]

Problem

Conformance compares audio by hashing samples in a canonical float32 representation. Every adapter reimplements the same conversion:

• Python: src/conformance/pcm.py FloatPcmHasher
• Rust: adapters/sendspin-rs/client/src/main.rs:85-133 FloatPcmHasher
• Swift: adapters/SendspinKit/client/Sources/ConformanceSendspinKitClient/main.swift:~152-231
• C++: adapters/sendspin-cpp/client/main.cpp (openssl SHA-256 + manual normalization)
• Go: server-side hashing lives in adapters/sendspin-go/internal/conformance/common.go

Each reimplementation is a place for subtle divergence (bit-depth normalization, 24-bit packed sample handling, endianness, sample ordering). It's also a symptom: SDKs expose raw bytes and leave "make these comparable across languages" to the harness authors.

Why this matters

• When a client fails PCM comparison, bisecting "is this a real playback bug or a hasher bug?" is hard when every language ships its own hasher.
• If the SDK shipped a canonical-samples iterator (or even a hash helper), adapters would stop needing this file.

Proposed SDK additions

Either:

1. A per-SDK canonical_samples(iter_of_bytes, bit_depth) -> iter_of_f32 helper (language-idiomatic) that matches the harness's Python reference byte-for-byte.
2. Or a shared protocol-side helper module (the time-filter repo precedent is cited in the spec itself) that every SDK can pull in.

Update all language adapters to use the helper instead of their hand-rolled hasher.