chore: no em-dashes in anything a reader sees

The GitHub About line had one, which is what prompted this. Fixing that made it
worth sweeping the rest.

All prose is clean: README, CONTRIBUTING, every doc, the justfile, the scripts,
the devcontainer files and the PR template.

All printed output is clean too, which mattered more than it sounds. Ninety-eight
em-dashes lived inside println strings, so they were in the first line of every
command. Banners became colons, the line-leading ones that were acting as bullets
became hyphens, and three that read badly as colons became commas or full stops.

Code comments were left alone. There are 97 of them across the core crates and
the kernel source, they predate this work, and they are the authors' own voice
rather than anything generated.

Verified by reading the output rather than trusting the substitution: probe,
gpu-selftest, model-decode, code-object-info and the OLMo gates all still read
correctly. 169 tests, the 79-command model API gate, the plugin contract, the
demo sandbox gate and all policy gates pass.

Author: maincode-dave

.devcontainer/Dockerfile

@@ -1,12 +1,12 @@
-# mainarch dev image — a clean Rust toolchain that talks to AMD silicon at the
+# mainarch dev image. A clean Rust toolchain that talks to AMD silicon at the
 # KERNEL ABI. Deliberately NO ROCm / HIP / HSA runtime: this project replaces
 # that stack rather than building on it. The only AMD-specific things needed are
 #   * the kernel UAPI headers (linux/kfd_ioctl.h) shipped by linux-libc-dev
 #   * libdrm's amdgpu_drm.h for the DRM render-node ioctls
 #   * LLVM/clang, whose upstream amdgpu backend can emit gfx950 ISA on its own
 # The GPU, when present, is reached purely via /dev/kfd + /dev/dri passthrough
 # (see .devcontainer/gpu/devcontainer.json). The image itself builds and runs
-# fine on a machine with no AMD GPU at all — every CPU-only path still works.
+# fine on a machine with no AMD GPU at all, and every CPU-only path still works.
 ARG BASE=docker.io/library/ubuntu:24.04
 FROM ${BASE}
 

.devcontainer/post-create.sh

@@ -11,7 +11,7 @@ echo
 if [ -e /dev/kfd ] && [ -d /sys/class/kfd/kfd/topology/nodes ]; then
   echo "AMD kernel driver visible: /dev/kfd is present."
   ls -l /dev/kfd /dev/dri/renderD* 2>/dev/null | head -4
-  echo "GPU lane available — 'just demo' will run the live KFD/AQL path."
+  echo "GPU lane available. 'just demo' will run the live KFD/AQL path."
 else
   echo "No /dev/kfd in this container: running the CPU-only lane."
   echo "Everything in 'just tour' still works; the live GPU proofs are skipped."

.github/pull_request_template.md

@@ -21,8 +21,8 @@ the evidence validates.
 - [ ] Focused model API checks, if `model_api` changed
       (`python3 tools/check_model_api_public_examples.py`)
 - [ ] Hardware gate, if the claim involves GPU execution
-- [ ] Apples-to-apples baseline, if the claim involves performance —
-      say which configuration the baseline ran in
+- [ ] Apples-to-apples baseline, if the claim involves performance.
+      Say which configuration the baseline ran in
 
 ## Evidence
 

CONTRIBUTING.md

@@ -14,9 +14,9 @@ and honestly described.
   is enforced by `just check-policy`, not just by review.
 - **Do not claim GPU execution, correctness, or performance without the run that
   backs it.** Paste the command and its output in the pull request. A
-  performance claim needs an apples-to-apples baseline — the relevant ROCm,
-  RCCL, Composable Kernel, vLLM, or SGLang result, or a prior in-repo number —
-  and needs to say which configuration the baseline ran in.
+  performance claim needs an apples-to-apples baseline, meaning the relevant
+  ROCm, RCCL, Composable Kernel, vLLM, or SGLang result, or a prior in-repo
+  number, and it needs to say which configuration that baseline ran in.
 - **Numerics go against a reference.** New kernels compare against an `f64` host
   implementation, with the tolerance stated.
 - **Keep CPU-only work labelled CPU-only.** The model API layer plans, validates,
@@ -29,9 +29,9 @@ and honestly described.
 
 Use the devcontainer.
 
-- `.devcontainer/devcontainer.json` — the default. CPU only, works on any
+- `.devcontainer/devcontainer.json` is the default. CPU only, works on any
   x86-64 Linux host, no GPU required.
-- `.devcontainer/gpu/devcontainer.json` — adds `/dev/kfd` and `/dev/dri`
+- `.devcontainer/gpu/devcontainer.json` adds `/dev/kfd` and `/dev/dri`
   passthrough. Use this on an AMD GPU host. The host user must be in the
   `render` group before the container starts.
 
@@ -111,7 +111,7 @@ spend a run on them.
 - The change is scoped to one clear atom.
 - Formatting, tests, and the gates appropriate to the change pass.
 - New behavior has evidence: the command, the hardware, and the output.
-- New claims state their negative scope — what the change does *not* prove.
+- New claims state their negative scope, meaning what the change does *not* prove.
 - CPU-only, preflight-only, or non-executing work is labelled as such.
 - Documentation and README status text match the actual behavior.
 - No unrelated local files, generated output, or secrets.

baseline/README.md

@@ -1,7 +1,7 @@
 # Baseline: upstream rccl-tests (reference only)
 
-`mainarch-collectives` mirrors the **rccl-tests methodology** — the same size
-sweep, correctness check, and algbw/busbw convention — so the numbers are
+`mainarch-collectives` mirrors the **rccl-tests methodology**, the same size
+sweep, correctness check, and algbw/busbw convention, so the numbers are
 directly comparable. This directory holds reference runs captured from the
 *existing* ROCm stack to compare against.
 
@@ -38,7 +38,8 @@ This runs both stacks over the same 8-GPU ladder and writes
 `baseline/allreduce-comparison.txt`.
 
 **Pick the RCCL configuration honestly.** `bench/compare-allreduce.sh` defaults
-to `RCCL_ARGS="-t 8 -g 1"` — 8 threads, one GPU each — which is RCCL's *fastest*
+to `RCCL_ARGS="-t 8 -g 1"`, meaning 8 threads with one GPU each, which is
+RCCL's *fastest*
 layout and therefore the right comparator. `-g 8` puts one host thread in charge
 of all 8 GPUs, which serializes RCCL's launch path and flatters mainarch. If you
 quote a `-g 8` number, say so.

baseline/allreduce-comparison.txt

@@ -1,4 +1,4 @@
-# all-reduce: mainarch (raw KFD/AQL, no ROCm) vs RCCL — 8x gfx950 (MI355X)
+# all-reduce: mainarch (raw KFD/AQL, no ROCm) vs RCCL, 8x gfx950 (MI355X)
 # generated by bench/compare-allreduce.sh
 # busbw convention: algbw * 2*(R-1)/R; lower time is better
 #

bench/compare-allreduce.sh

@@ -6,7 +6,7 @@
 # container built from upstream rccl-tests; mainarch runs natively from this
 # workspace.
 #
-# You must supply the RCCL image yourself — there is no public image with
+# You must supply the RCCL image yourself. There is no public image with
 # rccl-tests prebuilt. See baseline/README.md for how to build one:
 #
 #   MAINARCH_RCCL_IMAGE=<your-rccl-tests-image> bench/compare-allreduce.sh
@@ -51,7 +51,7 @@ docker run --rm --device /dev/kfd --device /dev/dri \
   2>/dev/null | grep -E "^ +[0-9]+ +[0-9]+ +float" > "$rh" || true
 
 {
-  echo "# all-reduce: mainarch (raw KFD/AQL, no ROCm) vs RCCL — 8x gfx950"
+  echo "# all-reduce: mainarch (raw KFD/AQL, no ROCm) vs RCCL, 8x gfx950"
   echo "# generated by bench/compare-allreduce.sh"
   echo "# busbw convention: algbw * 2*(R-1)/R; lower time is better"
   echo "# RCCL invocation: all_reduce_perf -b $MINB -e $MAXB -f 2 $RCCL_ARGS"

bench/compare-rccl.sh

@@ -1,5 +1,5 @@
 #!/usr/bin/env bash
-# compare-rccl.sh — side-by-side rccl all-reduce: ROCm baseline vs mainarch.
+# compare-rccl.sh: side-by-side rccl all-reduce, ROCm baseline vs mainarch.
 #
 # THE PATTERN (see README / docs/ROADMAP.md):
 #   Run this ON THE HOST with access to the container engine.
@@ -225,7 +225,7 @@ echo ">> ROCm baseline  (${ROCM_IMAGE:-<unset>})"
 BASELINE_OK=0
 MAINARCH_COMPARE=0
 if [ -z "$ROCM_IMAGE" ]; then
-  echo "   SKIPPED — set ROCM_IMAGE to a ROCm image with rccl-tests built."
+  echo "   SKIPPED. Set ROCM_IMAGE to a ROCm image with rccl-tests built."
   echo "   See baseline/README.md to build one (ROCm + RCCL + rccl-tests)."
 else
   if [ "$DRY_RUN" != "1" ] && ! has_image "$ROCM_IMAGE"; then
@@ -334,7 +334,7 @@ echo "mainarch -> $MAINARCH_OUT"
       fi
     fi
   else
-    echo "  STATUS: both outputs not present yet — set ROCM_IMAGE and rerun for a true apples-to-apples ratio."
+    echo "  STATUS: both outputs not present yet. Set ROCM_IMAGE and rerun for a true apples-to-apples ratio."
   fi
 fi
 

crates/mainarch-cli/src/demo.rs

@@ -1430,7 +1430,7 @@ pub fn serve(bind: &str, node: u32, olmo: Option<Arc<crate::olmo_lane::OlmoLane>
     match &olmo {
         Some(lane) => {
             println!(
-                "  OLMo 2 lane: {} resident on node {} — {} layers, hidden {}, vocab {}, {:.2} GiB",
+                "  OLMo 2 lane: {} resident on node {}, {} layers, hidden {}, vocab {}, {:.2} GiB",
                 lane.model_id,
                 lane.node,
                 lane.layers,