Add L1 schema plan and L2 specialized writer

[intech]

Summary

Replaces reflective encoding with a compiled-plan interpreter on the
toBinaryFast entry point. Each DescMessage is compiled once into a
flat Int32Array opcode stream plus side tables (field names,
pre-encoded tag bytes, sub-plans, map-entry plans, oneof case tables);
a single dense switch in executeSchemaPlan walks the stream with
every BinaryWriter call inlined so V8 keeps monomorphic receivers
on the hot path.

Implementation follows the 20 pinned decisions in the L1/L2 design
spec (analysis/p1-t4-l1-l2-design-spec.md, P1-P20).

Approach

• L1: flat Int32Array opcodes per schema, variable stride (2 for
  singular scalars, 3 for message/list/map/oneof), side-tables indexed
  by slot. Compiled once, cached in WeakMap<DescMessage, SchemaPlan | null>. Cycle-safe two-phase compile.
• L2: specialized writers inlined into the interpreter switch.
  ASCII fast path and int64 tri-dispatch inherited from L0 (no
  duplication, P11/P12). Pre-encoded tag bytes emitted via
  writer.raw(tags[slot]) (P13). Packed repeated scalars use a tight
  inline loop inside a single fork/join (P14). Element dispatch
  factored through a small writeScalarByOp helper so V8 sees one
  call site per writer method (P19).
• Fallback: proto2 groups, delimited-encoded messages inside
  lists, and messages carrying unknown fields with
  writeUnknownFields: true all transparently delegate to the
  reflective toBinary (P5, P10).

Measurements

Node.js 25.8.1, 2-second windows, warmup + 50-iter outer loop.

Workload	Reflective	toBinaryFast (L1+L2)	Delta
PerfMessage (nested, 16,132 B)	8,961 ops/s	15,861 ops/s	+77.0%
ScalarValuesMessage (97 B)	453,597 ops/s	505,119 ops/s	+11.4%
MessageFieldMessage (9 B)	737,761 ops/s	725,993 ops/s	-1.6% (noise)
PerfMessage heapUsed/op	664 B	119 B	-82%

Byte-parity verified against reflective on all three fixtures and on
the 14-case parity test suite.

Note: the external benchmarks/ folder used for the L0 OTel
100-span measurements is not merged into the L0 branch. The
PerfMessage fixture (100 nested message payload bundling lists,
maps, oneofs, and strings) is the closest representative workload
available in-tree.

Gates

• Byte-parity on all fixtures (14 new parity tests + PerfMessage round-trip)
• 2,889 existing tests pass (2,875 pre-existing + 14 new)
• Throughput improvement on realistic workload (+77%)
• Memory improvement on realistic workload (-82%)
• No regression on SimpleMessage-style workload (within noise)
• Diff under 1,500 LOC (971 LOC in src/, 192 LOC tests)

Stacked on

PR #8 (L0 contiguous-buffer writer). Cannot merge until L0 merges;
rebases automatically once L0 lands on main.

🤖 Generated with Claude Code