protobuf: unrevert and fix hyperpb processor

[mmatczuk]

• Revert "protobuf: remove hyperpb"
• protobuf: fix leak in hyperpb parser
• protobuf: guard against zero RecompileInterval in hyperpb decoder

[claude]

Bug: Integer division by zero — if RecompileInterval is 0 (the zero value of int64), the modulo operation p.seen.Add(1) % p.opts.RecompileInterval will panic at runtime.

Since ProfilingOptions is an exported struct with no validation, and NewHyperPbDecoder always creates a non-nil profile (line 30), the state.profile != nil guard won't protect against this — the first call to WithDecoded will panic if a zero-value ProfilingOptions{} is passed.

Add a guard (e.g., p.opts.RecompileInterval > 0 &&) before the modulo, or validate RecompileInterval in NewHyperPbDecoder.

[claude]

Commits
LGTM

Review
Reverts the hyperpb removal and fixes a memory leak by calling shared.Free() before returning Shared objects to the pool. The concurrent state-swap logic for profile-guided recompilation is sound.

1. Division by zero in hyperPbParser.WithDecoded — decode_hyperpb.go:79: p.seen.Add(1) % p.opts.RecompileInterval panics when RecompileInterval is 0 (the zero value of int64). Since ProfilingOptions is exported with no validation and NewHyperPbDecoder always creates a non-nil profile, passing a zero-value struct causes an immediate panic on the first message.

[claude]

The literal values 0.01 and 100_000 for profiling rate and recompile interval are repeated in three locations across two files (serde_protobuf.go, processor_protobuf.go x2) without named constants. Per the project's Magic Numbers code style rule:

Name all numeric constants. Every literal number in logic must have a clear meaning through a named constant or variable.

Define these as named constants (or a package-level default ProfilingOptions) in common and reference them from all three call sites. This also reduces the risk of the values diverging across locations.

[claude]

Commits
LGTM

Review
This PR re-introduces hyperpb as a profile-guided optimized protobuf decoder with platform-aware fallback, fixes a pool-related memory leak, and guards against division by zero on RecompileInterval. The concurrency design (atomic state swap, CAS-gated background recompile) is sound.

1. Magic numbers for profiling options (serde_protobuf.go#L88-L93): The literal values 0.01 and 100_000 are repeated across three call sites in two files without named constants, violating the project's Magic Numbers rule. Define a shared default in common and reference it from all callers.

[claude]

Commits
LGTM

Review
This PR re-introduces the hyperpb decoder with proper fixes: a memory leak fix via shared.Free() before pool return, a guard against zero RecompileInterval to prevent modulo-by-zero panics, and extraction of DefaultProfilingOptions to eliminate magic literals. The concurrency design is solid — atomic pointer swaps for state, CAS-guarded recompilation, and pool-per-state to ensure old pools drain cleanly with the old state. Platform fallback via complementary build tags (arm64 || amd64 vs \!(arm64 || amd64)) is correct. All callers of the changed NewDynamicPbDecoder signature have been updated.

LGTM

[claude]

Missing error wrapping: The unmarshal error is returned bare, but the existing dynamicPbParser.WithDecoded wraps it with message-type context:

return fmt.Errorf("unmarshalling protobuf message: '%v': %w", p.msgType.Descriptor().FullName(), err)

Since ProtobufDecoder implementations are interchangeable, error messages should be consistent. Wrap the error here with equivalent context (message type full name) so callers get the same diagnostic information regardless of which decoder is active. This also aligns with the project error handling pattern requiring fmt.Errorf wrapping with %w.

[claude]

Commits
LGTM

Review
This PR reintroduces the hyperpb decoder with a profile-guided recompilation strategy, a fallback to dynamicpb on unsupported platforms, and benchmarks comparing both decoders. The concurrency design (atomic state swaps, pool-per-state for GC-friendly lifecycle) is solid.

1. Missing error wrapping in hyperPbParser.WithDecoded (decode_hyperpb.go#L76-L77): The unmarshal error is returned bare (return err), but the existing dynamicPbParser wraps with message-type context. Since the two implementations are interchangeable behind the ProtobufDecoder interface, error messages should be consistent. This also violates the project error handling pattern requiring fmt.Errorf wrapping with %w.

[claude]

Commits
LGTM

Review
This PR re-introduces the hyperpb decoder with a fix for the pool/GC leak and adds profile-guided optimization (PGO) support with configurable profiling options. The concurrency design is sound: atomic state swaps with CAS for exclusive recompilation, pool-per-state isolation for safe GC, and a guard against zero RecompileInterval. The fallback on unsupported platforms gracefully degrades to dynamicpb.

LGTM