Release prep for v0.5.0.

What's in this PR

• Version bump — workspace 0.4.0 → 0.5.0, all intra-workspace dep pins, README.md / docs/guide.md / docs/migration-from-{prost,protobuf}.md / buffa/src/lib.rs version refs, gh release download examples in the guide.
• Cargo.lock regen — workspace, all 4 examples (addressbook, conflicts, envelope, logging), and both benchmark crates.
• CHANGELOG — ## [Unreleased] → ## [0.5.0] - 2026-05-05 with:
  • a top-of-section blurb stating that consumers with checked-in generated code must regenerate (0.5.0 codegen output references ViewReborrow, decode_bytes_to_bytes, and __private::arbitrary_bytes, none of which exist in buffa 0.4.0);
  • a ### Breaking changes section for #[non_exhaustive] on GeneratedFileKind;
  • footer compare-links updated (and a missing [0.4.0] link added).
• Benchmark refresh — re-ran task bench-cross and task bench-charts on the same hardware as the prior baseline (Intel Xeon Platinum 8488C). README tables and all 25 SVG charts updated; 5 new build-encode-*.svg charts added. The Binary encode table picks up the buffa (view) and prost (bytes) columns the chart generator now emits, and the README gains a Build + binary encode section that was previously omitted.

What's in 0.5.0

15 commits since v0.4.0 (f444eaba..ffcd6c08):

Why 0.5.0, not 0.4.1

#[non_exhaustive] on buffa_codegen::GeneratedFileKind plus the new Companion variant (#91) are SemVer-breaking for any consumer with an exhaustive match. The known consumer (connect-rust) only uses == and is unaffected, but the contract changed. Several other PRs change codegen output in ways that couple codegen and runtime versions (#84, #90, #95), which is also a minor-bump signal.

Benchmark deltas vs prior README baseline

The prior baseline was committed on 2026-04-22 (a0c668c0, pre-v0.4.0), so the comparison spans the 0.4.0 and 0.5.0 changes. Run-to-run Docker noise is ±5–15%; the largest deltas were investigated with a native bench (±1%) and perf before opening this PR.

• GoogleMessage1 decode: 1,014 → 905 MiB/s (−11%) in Docker. Native bench shows no v0.4.0 → v0.5.0 change (both 221.2-221.3 ns / 983 MiB/s). There is a real ~2% delta vs the prior baseline (216.8 ns / 1003 MiB/s), but it shipped in v0.4.0. perf record attributes it to drop_in_place::<GoogleMessage1> doubling (1.21% → 2.37% of cycles) plus Message::decode no longer being inlined into the benchmark closure. Bisect lands the change at Externalize size cache: remove __buffa_cached_size from generated structs #22 (size-cache externalization), which removed __buffa_cached_size: CachedSize from generated structs; GoogleMessage1SubMessage shrank from 160 → 152 bytes, crossing a glibc malloc size-class boundary. A small, intentional layout cost from making generated structs Send/Sync. The Docker −11% headline is ~5× the actual effect.
• AnalyticsEvent encode: 656 → 582 MiB/s (−11%) and GoogleMessage1: 2,594 → 2,441 (−6%). Likely attributable to the write_to field-number-order change in codegen: emit write_to fields in field-number order, not by kind #78 (a 0.4.0-era PR) — the previous group-by-kind ordering had cache benefits for messages with mixed field cardinality, traded for cross-implementation byte-equivalence. Intentional, documented trade-off (CHANGELOG codegen: write_to emits singular fields before repeated fields, not in field-number order #75).

Other deltas are in the noise band or improvements (MediaFrame decode +5%, LogRecord JSON decode +5%). No regressions in this release — the deltas vs the README's prior numbers shipped in v0.4.0.

Pre-tag verification (matches CI)

• cargo test --workspace — 1571 pass
• cargo clippy --workspace --all-features --all-targets -- -D warnings — clean
• RUSTDOCFLAGS=-D warnings cargo doc --workspace --all-features --no-deps — clean
• cargo +1.95 fmt --all --check — clean
• task gen-wkt-types — no drift
• task lint-md — clean

After merge

• Tag v0.5.0 on the merge commit
• Dispatch publish-crates.yml
• Bump connect-rust to depend on buffa 0.5.0, adopt apply_companions (replaces the kind: GeneratedFileKind::Owned workaround in connectrpc-codegen/src/codegen.rs), then cut connect-rust 0.4.0