One bench, one corpus: retire the §1.2 example-corpus rows from every runner

[gafferongames]

Retires the §1.2 example-corpus bench rows — rigidbody_moving,
rigidbody_at_rest, chat, test, inputpacket, shipcreate,
probe_header, probebits, probearray, testdata and real_packet —
from the c, cpp, go, rust, cs and js runners, together with every
hand-written pin, vary function, field check and per-shape driver that fed
them. They were the last hand-coded shape measurement code in the harness.

The owner ruling reaches them directly: "there should be only a single
schema bench: Bench.schema, it generates per-language stuff that is how we
measure how efficient serialize and schema is per-language" / "We do all
profiling inside schema now." #199 removed the hand-coded Bench-corpus
shapes and family rt; this removes the second corpus that rode the full
sweep beside them. BenchMixed is now the one measured shape in every one
of the nine runners.

real_packet rides along even though its shape lives in
bench/corpus/RealWorld.schema rather than examples/: it is a second
bench schema with its own hand-written pin and vary body, which is exactly
what "a single schema bench: Bench.schema" forbids.

What is deleted, and what is kept

Deleted: measurement code only — 480 lines from bench/c/bench_main.c, 535
from bench/cpp/bench_main.cpp, and the equivalents in go, rust, cs and js;
the retired rows' lines in bench/inline-budget.txt (a budget line for a leg
that produces no verdict fails the gate as a vanished verdict); the
corpus list in bench/tools/relative.go; the now-dead example /
realworldcorpus dependencies in go.mod, Cargo.toml and the C/C++
include paths.

Kept, because make test proves they are load-bearing:
examples/*.schema, bench/corpus/RealWorld.schema, all of generated/
and every testdata/wire/*.bin. The cross-language conformance suite pins
these instances byte-for-byte — test/main.cpp, test/c/main.c and the go
/ rust / cs / js / java / dart / elixir port suites for the example
goldens; test/bench/main.cpp, test/bench/c_main.c and the java / js /
dart / elixir suites for real_packet. Nothing under examples/,
generated/ or testdata/ is touched by this PR. The two generated units
that no longer have a bench consumer keep their compile gates:
generated/bench/rust-realworld its own cargo build in the Makefile, and
generated/bench/cs/realworld a Compile item in schemabench.csproj
documented as a compile gate with no bench reader (issue #80's lesson —
that unit once shipped uncompilable while every other gate stayed green).

Family bits (§1.4) is out of scope and untouched.

Also, both ruled out of #199's report: §1.8's SUSPENDED bench_wstring
definition is deleted. schema defers wstring (SPEC §4.10) and #188 owns
growing the language; a suspended row for a construct the language cannot
express is fossil, not a record. When schema expresses wstring the row is
defined then, against what the language actually emits. (#199's other ruled
item — bench/LOCK's control 2 naming rows that no longer exist — is a
separate PR touching only that file, per LOCK's own amendment protocol.)

Verification

• make test green, 26 conformance legs, nine languages. Nothing
  deleted here was load-bearing for a test; the fixtures that were, stayed.
• All nine runners build and run bench_mixed with their §1.5 golden
  gates green: c, cpp, go, rust, cs, js, java, dart, elixir.
• corpus_id 6b213fbfa1a03a99 on all nine --quick legs, unchanged.
• bench/run.sh --quick prints the two-column table: c 98%, cpp 100%,
  rust 153%, java 156%, cs 360%, dart 392%, js 469%, go 581%, elixir 3957%.
• bench/tools/inline-gate.sh selftest passes all 25 fixtures;
  inline-gate.sh leg {c,cpp,go,rust,cs} all PASS, 0 unguarded.
• Emitters and generated/ byte-unchanged: a full make from clean
  leaves the working tree empty. This PR's diff touches bench/ only.

Ledger — and yes, this is another #194 sub-era boundary

Receipts and the full write-up: bench/results/exrows-air/.

before → after → before for the c and cpp legs in one sitting, then an
A/B/A/B reproduction sitting per language (a macOS media-indexing daemon
held a core and put sitting 1's cpp control over §2.3's 15% spread
threshold; the failed-control CSVs are landed rather than dropped).

leg	row	before	after	control drift
c	bench_mixed write	6.43 M/s	6.90 M/s	+7.4% (control +0.1%)
c	bench_mixed round_trip	3.48	3.43	-1.3% (control -0.9%)
c	bitpacker write/read	—	—	flat, ≤0.4%
cpp	bench_mixed write	7.07	7.08	+0.2%
cpp	bench_mixed round_trip	3.46	3.37	-2.8%
cpp	bitpacker read	87.4 k/s	84.9 k/s	-2.8%

bench_bitpacker is byte-identical in both trees in both languages —
the diff moves only its call site's position in the file — and cpp's still
moves 2.8%. That is the negative control. C's +7% reproduces in two
independent windows against controls under 1%.

This commit is a #194 absolute-ledger SUB-ERA BOUNDARY for the c and cpp
legs. Removing ~500 lines from each TU moves layout again, exactly as #199
found one merge earlier. Do not read C's +7% as an improvement or cpp's
-2.8% as a regression; do not smooth it and do not chase it.

The full c/cpp sweep's corpus_id moves a4bc52bae5343e60 →
6b213fbfa1a03a99 — required, since the leg no longer loads the eleven
retired goldens, so the full sweep's id now equals --quick's. §5.3 rule 2
mechanically refuses every ratio across the boundary, which is the loud
re-pricing §1.7 rule 3 demands. The goldens themselves are byte-unchanged.

Note for PR #200

This PR deletes exactly the hand-coded shape measurement code that #200's
shape gate exists to keep out, so that gate's ledger counts will change
under it. bench/tools/shapegate and SHAPE-GATE.allow are #200's and are
not touched here.

🤖 Generated with Claude Code