refactor: replace unmaintained bincode with postcard + borsh

Summary

Replace the unmaintained bincode crate (RUSTSEC-2025-0141) with two actively maintained serialization backends behind feature gates.

Changes

Dependencies

• Removed: bincode = "1" (workspace)
• Added: postcard = "1" (serde-native, community standard replacement)
• Added: borsh = { version = "1", features = ["derive"] } (NEAR Protocol, high performance)

Feature Gates

Modified Files

• Cargo.toml — workspace deps
• foyer-common/Cargo.toml — features and dependencies
• foyer-common/src/code.rs — three mutually exclusive Code impl paths with #[cfg]
• foyer-common/src/error.rs — bincode_error() -> postcard_error()
• foyer/Cargo.toml — add borsh pass-through feature
• examples/Cargo.toml — add borsh pass-through feature

New Benchmark

foyer-common/benches/bench_serde/ rewritten to directly compare postcard, borsh, and manual encoding:

• postcard.rs — postcard backend (#[cfg(feature = "serde")])
• borsh.rs — borsh backend (#[cfg(feature = "borsh")])
• manual.rs — hand-written LE encoding (always available)

Benchmark Results

Test machine: Linux 6.8, 100 samples per benchmark, 3s warmup. Test data: struct Entry { id: u64, label: String, payload: Vec<u8> }

Encode

Decode

Encoded Size (raw = id + label + payload)

Key Findings

1. borsh stays close to manual — only 1.5-2x slower than raw memcpy for both encode and decode across all sizes
2. postcard is significantly slower for large payloads: 21x slower encode and 34x slower decode than manual at 4 MiB. The varint encoding overhead dominates for bulk data.
3. At 64 MiB the gap narrows as memory bandwidth becomes the bottleneck — borsh ~5x, postcard ~5.4x vs manual for encode
4. All three produce nearly identical encoded sizes — only 5-16 bytes of metadata overhead regardless of payload size
5. manual encoding is always fastest but requires handwritten encode/decode — the baseline cost is essentially memcpy

Migration Notes

• Wire format change: postcard and bincode produce different byte representations. Disk caches written with bincode cannot be read with postcard. Users with persistent disk caches must clear or migrate them.
• Error::bincode_error() has been replaced with Error::postcard_error() (available when the serde feature is active and borsh is not).
• Users who derived Serialize + DeserializeOwned for Code continue to work with the serde feature, no code changes needed.
• Users who want maximum performance should consider the borsh feature and derive BorshSerialize + BorshDeserialize.

---

Generated with Claude Code