[jasperpotts]

What and why

This PR applies the straight-line approach from #995 to the other two hot paths (docs/fast-codegen-plan.md section 6.1, items 1-3 and 6):

Performance (AMD EPYC 9275F): toBytes of a freshly built block item is 2.1× faster (968 → 452 ns), of a signed transaction 1.7× (96 → 58 ns), and whole-block stream writes are 15-23% faster. Step 1 as a whole (#994-#996) makes block item array writes 28-30% faster. Tier A, on by default.

• protobufSize() in the model. It now uses literal tag sizes and inline default checks, and measures each nested message through that model's own protobufSize(). That is a direct call on the model type. Before, it went through Codec.measureRecord on the shared, megamorphic call site. Every block item that is built and written once is measured once, so this is part of every fresh write.
• The stream writer write(T, WritableSequentialData). It writes constant tags with writeByte / writeByte2..4 and checks defaults inline. Each codec gets a new public static void writeTo(T, WritableSequentialData), which nested messages from the same build call directly.
• Codec.toBytes now writes through write(T, byte[], 0), the array writer, instead of the stream writer wrapped around a BufferedData.
• One shared generator. InlineFieldCodeGenerator produces the inline code for all three targets: byte array, stream and size. The field order, default checks, oneof switches and fallbacks are written once. The byte-array writer from perf: generate straight-line byte array writers with direct nested calls #995 moves into it unchanged.

ProtoWriterTools.sizeOfUnsignedVarInt64 and sizeOfStringNoTag become public, for generated models. New writeStringNoTag and writeBytesNoTag stream helpers are added.

Generated code: TransactionID.protobufSize() before and after

Before:

// [1] - transactionValidStart
_size += sizeOfMessage(TransactionIDSchema.TRANSACTION_VALID_START, transactionValidStart, com.hedera.hapi.node.base.Timestamp.PROTOBUF);
// [4] - nonce
_size += sizeOfInteger(TransactionIDSchema.NONCE, nonce, true);

After:

// [1] - transactionValidStart
final var transactionValidStart = transactionValidStart();
if (transactionValidStart != null) {
    final int size = transactionValidStart.protobufSize();
    _size += 1 + sizeOfVarInt32(size) + size;
}
// [4] - nonce
final var nonce = nonce();
if (nonce != 0) {
    _size += 1 + sizeOfVarInt32(nonce);
}

The stream writer, for the same message:

final var transactionValidStart = data.transactionValidStart();
if (transactionValidStart != null) {
    out.writeByte((byte) 0x0A); // tag: field 1, wire type 2
    final int size = transactionValidStart.protobufSize();
    out.writeVarInt(size, false);
    com.hedera.hapi.node.base.codec.TimestampProtoCodec.writeTo(transactionValidStart, out);
}

Tests

• ByteArrayWriterGenerationTest is extended: the stream writer and protobufSize() use direct calls for messages from this build, and the Codec / sizeOfMessage helpers for dependencies.
• pbj-core: spotlessApply qualityGate build passes.
• pbj-integration-tests: ./gradlew build and -PpbjPerformance=max both pass, 326,323 tests each. The protoc comparisons check measureRecord, stream writes and toBytes for every test object.

Performance: AMD EPYC 9275F, before = #995's branch, after = this PR

3 forks, 3 × 1 s warmup, 5 × 1 s measurement, -prof gc.

Benchmark	Before ns/op	After ns/op	Change	Before B/op	After B/op
BlockItemBench.itemToBytesCached	1,266 ± 89.7	935.2 ± 37.0	-26.2%	419	339
BlockItemBench.itemToBytesFresh	969.8 ± 48.3	451.8 ± 16.3	-53.4%	2,142	2,062
BlockItemBench.itemWriteByteArrayCached	296.4 ± 11.1	274.0 ± 4.7	-7.6%	0	0
BlockItemBench.itemWriteByteArrayFresh	483.3 ± 12.6	420.7 ± 27.0	-12.9%	1,723	1,723
BlockItemBench.transactionToBytesFresh	93.9 ± 3.7	57.9 ± 1.8	-38.3%	752	672
BlockItemBench.transactionWriteByteArrayCached	31.2 ± 1.4	32.4 ± 0.7	+3.6%	0	0
BlockItemBench.transactionWriteByteArrayFresh	47.4 ± 0.9	48.2 ± 2.2	+1.6%	399	399
SampleBlockBench.writePbjByteArray	57,289,139 ± 1,390,887	45,463,554 ± 830,612	-20.6%	385	313
SampleBlockBench.writePbjByteBuffer	57,514,555 ± 1,235,217	44,365,160 ± 1,025,929	-22.9%	392	295
SampleBlockBench.writePbjByteDirect	40,065,832 ± 1,218,275	34,305,294 ± 654,058	-14.4%	279	236
SampleBlockBench.writePbjOutputStream	67,003,336 ± 1,202,114	64,092,798 ± 2,631,824	-4.3%	19,605,947	19,605,926
SampleBlockBench.writePbjToBytes	75,216,791 ± 2,135,534	56,885,573 ± 2,100,181	-24.4%	19,606,065	19,605,870

• toBytes of a freshly built block item is 2.1× faster (968 → 452 ns), and 1.7× for a signed transaction (96 → 58 ns), with 80 B/op less allocated. This is the measuring pass plus routing toBytes to the array writer.
• Whole-block stream writes are 15-23% faster: SampleBlockBench into heap and direct BufferedData, and toBytes. Writing to an OutputStream is unchanged (-4%, within noise), because the stream itself dominates.
• write(T, byte[]) for fresh block items is 13% faster, because measuring is part of it. Cached-size writes and signed transactions are unchanged within noise.

Step 1 as a whole (#994 → this PR), BlockItemBench per block item:

	Before step 1	After	Change
write(byte[]), freshly built	602 ns	421 ns	-30%
write(byte[]), sizes cached	379 ns	274 ns	-28%
toBytes, freshly built	968 ns	452 ns	-53% (2.1×)
signed transaction toBytes	96 ns	58 ns	-40%

The plan expected about 2× for writing deeply nested messages. Step 1 reaches it for toBytes of fresh block items, and reaches 1.4× for the array write itself.

Generated code size (full HAPI build)

	main	#995	This PR
All classes	57,723,772	58,317,471	58,478,383 (+1.3% vs main)
main classes	45,502,478	46,042,637	46,203,717 (+1.5%)
*ProtoCodec classes	7,140,244	7,680,403	7,852,023 (+10.0%)

The plan expected step 1 to shrink generated code; it grows slightly, well within the 15% tier A limit.

Stacked on #995.

Fixes #990

🤖 Generated with Claude Code

[github-actions]

JUnit Test Report

   551 files  ±0     551 suites  ±0   52s ⏱️ -1s
 2 052 tests ±0   2 048 ✅ ±0   4 💤 ±0  0 ❌ ±0 
11 584 runs  ±0  11 556 ✅ ±0  28 💤 ±0  0 ❌ ±0 

Results for commit 0949dcf. ± Comparison against base commit 890bc4f.

♻️ This comment has been updated with latest results.

[github-actions]

Integration Test Report

    623 files  ±0      623 suites  ±0   47m 20s ⏱️ + 5m 49s
163 876 tests ±0  163 876 ✅ ±0  0 💤 ±0  0 ❌ ±0 
326 467 runs  ±0  326 467 ✅ ±0  0 💤 ±0  0 ❌ ±0 

Results for commit 0949dcf. ± Comparison against base commit 890bc4f.

♻️ This comment has been updated with latest results.

[jasperpotts]

Review follow-ups

• FIX: dead code. CodecMeasureRecordMethodGenerator.generateMeasureMethod no longer computes size lines for the placeholder the template no longer has. The helpers stay, for map entries. (This predates the series.)
• Merged in from perf: generate straight-line byte array writers with direct nested calls #995: the unique local names now apply here too, in InlineFieldCodeGenerator, so the stream writer and protobufSize() are covered.
• Thanks for checking the shared generator and the SINT32 size formula. Generating all three targets from one place was the aim, so that size and writer can't disagree.

[anthony-swirldslabs]

Many review comments from #995 apply to this fix, including the newly introduced class here, because they iterate on the same code.

[anthony-swirldslabs]

Please don't generate unbraced multi-line if's. Removing the \n might be the nicest fix.