Codegen: SipHash dominates CodeGenContext's internal naming/comment maps #396

The problem

CodeGenContext's internal bookkeeping — package/message name deconfliction, doc-comment lookup, oneof/message indexing (context.rs, comments.rs) — is built fresh from the whole descriptor set on every generate() call, using std::collections::HashMap. That defaults to SipHash, which is designed to resist hash-flooding from untrusted input.

This data only ever comes from trusted, locally-generated .proto descriptors (protoc/buf output), never attacker-controlled input, so SipHash's DoS resistance buys nothing here, and its cost is measurable.

Measured

Sampling (sample, macOS) a real ~2900-crate, ~2900-package codegen run against a large IDL corpus, sorted by top-of-stack leaf frame:

core::hash::sip::Hasher::write            6338 samples
hashbrown::raw::RawIterRange::fold_impl   1438
core::hash::BuildHasher::hash_one         1034 (+295 +54 +11 elsewhere)
HashMap::insert (several monomorphizations)  ~1067 combined
RawTable::reserve_rehash                   ~639 combined

Hashing-related frames were roughly 17% of all samples — the single largest cost bucket in the profile, ahead of any actual codegen logic (token generation, prettyplease formatting, etc.).

Fix

Swap context.rs and comments.rs's internal maps to a non-cryptographic hasher. foldhash is already resolved transitively via hashbrown, so this is a direct dependency on something already in the tree, not new third-party code.

oneof.rs's maps were tried too but reverted: they leak through public signatures in message.rs/view.rs/lazy_view.rs, too broad a blast radius for what should be a purely internal, output-invariant change. Left as std::HashMap for now — a future PR could thread a type alias through those signatures if it's worth the diff.

Verification

Output-invariant by construction (the change only affects which hash function internal maps use, never what gets generated) — confirmed with a full hash-diff of a real 2895-crate generation run, before and after: byte-for-byte identical. All existing buffa-codegen tests pass unmodified.

PR to follow.