Add QBS versus generated C benchmark

[ip332]

Summary

Add a same-schema benchmark comparing generated C with QBS-driven generic C. Both paths use benchmarks/schemas/workload.brd; QBS is generated at build time through quarry-schema-compiler --emit-qbs, and one initialized logical workload feeds both adapters. Correctness gates run before timing.

Correctness

• Deterministic QBS generation: PASS
• QBS image: 881 bytes
• BRF payload: 252 bytes
• Generated-C/generic-C BRF byte equality: PASS
• Generated checksum: 711237
• Generic checksum: 711237
• Public get_value() -> ARRAY -> record_array_get() path: PASS

Performance baseline

Metric	Generated C	QBS-driven C
Encode	459.7 ns/op	1,472.7 ns/op
Encode throughput	2.18M ops/s	679K ops/s
Post-validation access	94.3 ns/op	265.0 ns/op
Access throughput	10.6M ops/s	3.77M ops/s

QBS parse: 355 ns/parse (2.82M parses/s). Ratios are approximately 3.2x for encode and 2.8x for post-validation access. These are optimized local measurements, not universal performance claims.

Schema-artifact footprint

Artifact	Size
QBS image	881 bytes
Generated headers	5,060 bytes
Generated C source	41,195 bytes
Root generated object	9,392 bytes
Imported generated object	4,072 bytes
Combined generated objects	13,464 bytes

The 15.3x object/QBS ratio compares schema-specific artifacts only; it is not a complete application-footprint ratio. Both approaches require runtime/support code.

Generated object details: root .text 8,013 bytes/read-only 376 bytes; imported .text 3,485 bytes/read-only 108 bytes. Source/header counts are build-artifact measurements, not executable footprint.

Memory and integration

Configured parser/validation workspace is 13,056 bytes; exact consumed minimum is not exposed by the public API. Shared BRF buffers are treated consistently. The QBS comparison is integrated into the common benchmark runner, and generated object sections are printed through llvm-size when available.

Validation

Optimized benchmark build/run, common runner, strict C99 benchmark build, BRF equality, checksum equality, complete CTest (52/52), and git diff --check all pass.

Compatibility

Benchmark-only change. No production compiler/runtime, API/ABI, BRF, QBS, Schema IR, or compatibility epoch changes.

Deferred work

Embedded footprint, exact incremental generic-runtime attribution, multi-schema scaling, and production performance optimization are deferred. This PR establishes the pre-optimization baseline.