fix: -march=native spec parity in CI workflows + Dockerfile

[birdnamoo]

PR Draft — fix: -march=native spec parity in CI workflows + Dockerfile

Target repository: ecosystem/benchmark-bmb (submodule)
Branch: fix/march-native-spec-parity
Base: main
Draft prepared by: Cycle 2746 (autonomous), parent repo HEAD 627dd3a1
Push status: NOT pushed (HUMAN trigger required per HANDOFF E.4)

Title (≤70 chars)

fix: -march=native spec parity in CI workflows + Dockerfile

Body

Summary

Adds -march=native to 5 compile sites in CI workflows and Dockerfile to
match the published benchmark specification (docs/BENCHMARK_MASTERPLAN.md:135),
which mandates clang 18.x (-O3 -march=native) for the C baseline.

Discovered during Cycle 2743 (parent repo) policy-vs-code gap audit.

Sites changed (+5 / -5)

File	Line	Compiler	Before	After
.github/workflows/benchmark-multiplatform.yml	96	clang (Unix)	clang -O3	clang -O3 -march=native
.github/workflows/benchmark-multiplatform.yml	132	gcc (MinGW Win)	gcc -O3	gcc -O3 -march=native
.github/workflows/benchmark.yml	71	clang	clang -O3	clang -O3 -march=native
.github/workflows/benchmark.yml	118	clang	clang -O3	clang -O3 -march=native
Dockerfile	71	clang	clang -O3	clang -O3 -march=native

Already-compliant sites (no change):

• .github/workflows/benchmark-multiplatform.yml:94 — gcc Unix
• Dockerfile:64 — gcc
• run_benchmarks.sh:77 — clang
• run_wsl_bench.sh:30,45 — clang

⚠️ Baseline change — CI history discontinuity

This is a baseline change. After merge:

• BMB-vs-C ratios in CI may shift nominally (slightly less favorable to BMB
  or roughly equal, depending on workload). -march=native enables target
  CPU vectorization (AVX2/AVX-512), which mostly helps numeric kernels.
• CI history prior to this PR is not directly comparable with runs after
  this PR. First CI run after merge should be stamped as the new baseline.
• External consumers referencing CI history (e.g., README links to specific
  workflow runs) should be aware of the discontinuity.

Rationale

The benchmark spec (docs/BENCHMARK_MASTERPLAN.md:135) explicitly states:

C baseline: clang 18.x (-O3 -march=native)

5 of the 10 audited compile sites had -march=native missing. This PR
brings them into spec parity. Without -march=native, the C baseline runs
with conservative ISA (typically x86-64-v1 on Ubuntu runners), making the
BMB-vs-C comparison favor BMB unfairly. The BMB compiler itself uses LLVM
opt -O2 which can target the host CPU; the C baseline should match.

Verification

Local diff (git diff on fix/march-native-spec-parity branch):

.github/workflows/benchmark-multiplatform.yml | 4 ++--
.github/workflows/benchmark.yml               | 4 ++--
Dockerfile                                    | 2 +-
3 files changed, 5 insertions(+), 5 deletions(-)

Only -march=native is added; no other semantics change.

Post-merge actions

1. First CI run on main after merge → stamp result JSON as new baseline.
2. Update parent repo (lang-bmb) submodule pointer to include this commit.
3. Cycle 2743 carry-forward closed.

Related

• Parent repo Cycle 2743 log: claudedocs/cycle-logs/cycle-2743.md
• Parent repo HANDOFF Sequence E (decision adopted Cycle 2745)

---

HUMAN dispatch instructions (Sequence E.4):

# Submodule push
cd ecosystem/benchmark-bmb
git push -u origin fix/march-native-spec-parity

# Create PR via gh
gh pr create \
  --repo <owner>/<benchmark-bmb-repo> \
  --base main \
  --head fix/march-native-spec-parity \
  --title "fix: -march=native spec parity in CI workflows + Dockerfile" \
  --body-file <path-to-this-file> \
  --draft

[birdnamoo]

Withdrawn; merging directly to main per maintainer decision.