build: cache C++ and CUDA compilation in container builds

[DuncanBetts]

Container image rebuilds currently discard all previous compiler work whenever
COPY . . changes. In the reproduction below, adding one harmless line to one
.cpp file and one .cu file causes all 278 Ninja steps to run again, even
though 261 of the 263 compiler inputs are unchanged.

Cause. The Dockerfile invokes the C, C++, and CUDA compilers directly and keeps
the build tree inside the invalidated image layer. There is therefore no
compiler-output cache available after a source change. The repository also has
no persistent containerized development-build path.

Fix. Three files:

file	change
Dockerfile	install ccache; split out a reusable build-env stage; configure CMake's C, C++, and CUDA compiler launchers; retain compiler outputs in a 20 GiB cache mount; print cumulative statistics
compose.yaml	add a build service with the source mounted read-only and named volumes for the Ninja build tree and ccache
README.md	document the iterative podman compose run --rm build workflow

The image build and Compose paths serve slightly different cases:

• Dockerfile builds start with a fresh Ninja tree after source-layer
  invalidation, but ccache restores unchanged compiler outputs.
• Compose builds retain both /build and /ccache, allowing Ninja to skip
  unchanged targets before ccache is consulted.

Verification. The before and after recordings run the same 100×30 terminal
script against the same checkout:

1. perform a cold image build;
2. add the same harmless static_assert to one C++ and one CUDA file;
3. rebuild the image;
4. restore the exact source tree.

result after the two-file edit	before	after
Ninja steps	278	278
cacheable compiler calls on the rebuild	unavailable	263
compiler-cache hits	0	261
compiler-cache misses	263	2

The after recording prints cumulative totals for both builds:
526 cacheable calls, 261 hits, and 265 misses. The 265 misses comprise
the initial cold build's 263 misses plus the two genuinely changed
translation units.

Reproduction script (ninfer-cache-demo.sh):

#!/bin/sh
set -eux

demo_label=$1
cd /home/duncan/git/prefix-cache-research/repos/ninfer

printf '\n=== 1. Build image ===\n'
podman build --no-cache --tag "localhost/ninfer-cache-demo:${demo_label}" .

printf '\n=== 2. Apply one C++ and one CUDA source edit ===\n'
git apply --verbose /tmp/ninfer-cache-probe.patch
git diff --stat

printf '\n=== 3. Rebuild image after the same two-file edit ===\n'
podman build --tag "localhost/ninfer-cache-demo:${demo_label}" .

printf '\n=== Restore the exact source tree ===\n'
git apply --reverse --verbose /tmp/ninfer-cache-probe.patch
git diff --exit-code -- src/artifact/typed_binding.cpp src/ops/launcher/residual_add.cu

Recording — before:

Recording — after:

Marker timestamps are relative to the start of each recording:

chapter	before	after	duration (before / after)
cold image build	00:00.002	00:00.003	02:26.133 / 02:31.940
edit C++ + CUDA	02:26.135	02:31.943	00:00.004 / 00:00.002
rebuild after edit	02:26.139	02:31.945	02:07.750 / 00:21.837
restore source	04:33.889	02:53.782	00:00.005 / 00:00.003

Relevant documentation:

• ccache manual
• CMake compiler launcher variables
• Docker build cache mounts
• Compose named volumes
• Podman Compose

Scope. Build ergonomics only. The runtime stage, installed runtime packages,
CMake targets, generated executables, and inference behavior are unchanged.
Python and ccache are present only in the build environment.

Changes made using AI, with careful steering and review by myself. Thanks for creating Ninfer Neroued, I hope you find these build time savings useful in future work, you can take it as a token of my appreciation :).

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e64041e07a

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".