Run every lint-class check inside Docker (closes #38)

Add a root Dockerfile.lint that carries the checks as build steps -- a
`lint` stage running `hugo --minify --printPathWarnings` and a
`fmt-check` stage running the prettier check -- and reduce script/lint
and script/fmt-check to building their stage. A successful build is a
clean check. There is no host path and deliberately no "am I already
inside a container?" branch, which would be a host lint path in
disguise.

The two stages share a `base` whose first four instructions are
byte-identical to the main Dockerfile's, so the expensive
`RUN script/bootstrap` layer that compiles the pinned Hugo from source
is a cache hit against the main image instead of a second build of the
same thing.

Resolve the resulting recursion by splitting the checks by where they
run, not with an escape hatch. `make check` runs script/lint, so the
main Dockerfile can no longer `RUN make check`: that would be
docker-in-docker inside a bare Alpine with no docker client and no
daemon socket, and script/cibuild is what CI runs on every push. The
main Dockerfile therefore runs `make test`, the production build, and
script/cibuild builds it and then calls script/lint and
script/fmt-check. CI still covers the production build, lint and the
format check, and it runs exactly what a developer runs.

script/fmt stays on the host because it rewrites the working tree,
which a container build cannot do. That makes it the authoritative
copy of the prettier version, scope and flags that the fmt-check stage
duplicates; both sides carry a keep-in-sync note. The duplication is
forced: any `RUN script/fmt-check` inside the image is the recursion
again.

Caching is waived for the checks in the shape this repo already
settled: `ARG CHECK_EPOCH` with no default, declared and guarded
separately in each stage because ARG does not cross a FROM, with the
value expanded into the checked command as well as the guard so
invalidation does not rest on BuildKit's treatment of an unreferenced
ARG. All four image-building entrypoints now generate and pass it --
script/cibuild, script/docker, script/lint, script/fmt-check.

Verified: two consecutive script/lint runs on an unchanged tree both
executed hugo for real, with script/bootstrap CACHED; a constant-epoch
counterfactual restored the false green (exit 0, lint layer CACHED, no
hugo output); an empty epoch failed closed on the guard; a broken
template failed the lint stage and an unformatted README failed the
fmt-check stage, both reverted and re-run clean; script/cibuild and
`make check` are green with all three checks demonstrably executing.

.gitea/workflows/deploy.yml

@@ -108,43 +108,3 @@ jobs:
               run: wrangler pages deploy public --project-name=lora-vegas --branch=${{ github.ref_name }}
               env:
                   CLOUDFLARE_API_TOKEN: ${{ secrets.CLOUDFLARE_API_TOKEN }}
-
-    notify:
-        runs-on: ubuntu-latest
-        needs: [build, deploy]
-        # Run whether the build and deploy passed or failed. The Notify step
-        # carries continue-on-error: true so a notification that cannot be sent
-        # leaves the deploy's result exactly as it was.
-        if: always()
-        container:
-            # node 20.20.2-bookworm, 2026-08-09. Same pinned image as the deploy
-            # job, chosen because it ships curl; no extra packages to install.
-            image: node@sha256:8f693eaa7e0a8e71560c9a82b55fd54c2ae920a2ba5d2cde28bac7d1c01c9ba5
-        steps:
-            # needs.deploy.result is "success" only when the deploy job actually
-            # published; anything else -- a failed build or a failed deploy -- is
-            # reported as a failure with a link to the run.
-            - name: Notify
-              # curl failing (ntfy unreachable or erroring, or NTFY_TOPIC
-              # absent) must not fail the run, so continue-on-error lives on
-              # this step and not on the job: on this Gitea a job-level flag
-              # still marks the job failed and reddens the commit.
-              continue-on-error: true
-              env:
-                  RESULT: ${{ needs.deploy.result }}
-                  NTFY_TOPIC: ${{ secrets.NTFY_TOPIC }}
-              run: |
-                  short="$(printf '%.7s' "$GITHUB_SHA")"
-                  run_url="$GITHUB_SERVER_URL/$GITHUB_REPOSITORY/actions/runs/$GITHUB_RUN_ID"
-                  if [ "$RESULT" = success ]; then
-                      curl -fsS --max-time 10 \
-                          -H "Title: lora.vegas deployed from $GITHUB_REF_NAME" \
-                          -d "$short deployed to https://lora.vegas/" \
-                          "https://ntfy.sh/$NTFY_TOPIC"
-                  else
-                      curl -fsS --max-time 10 \
-                          -H "Title: lora.vegas deploy from $GITHUB_REF_NAME failed" \
-                          -H "Priority: high" \
-                          -d "Run: $run_url" \
-                          "https://ntfy.sh/$NTFY_TOPIC"
-                  fi

Dockerfile

@@ -1,16 +1,18 @@
-# Hugo static-site build image. The build runs the individual non-lint
-# checks -- script/test, a clean `hugo --minify` production build, and
-# script/fmt-check, the read-only prettier check -- so the image build
-# fails on any template, content, config or formatting error.
+# Hugo static-site build image. The build runs `make test` -- a clean
+# `hugo --minify` production build -- so the image build fails on any
+# template, content or config error.
 #
-# It deliberately does NOT run `make check`, and only the lint is
-# missing from what it does run. `make check` calls script/lint, and
-# script/lint is a `docker build` of Dockerfile.lint, so `RUN make
-# check` here would attempt a docker build inside a build step, in a
-# bare alpine with no docker client and no daemon socket. Putting
-# `make check` (or a `make lint`) back reintroduces exactly that
-# recursion. The lint is not skipped: script/cibuild runs script/lint
-# first, in its own container, before this build starts.
+# It deliberately does NOT run `make check`. `make check` runs
+# script/lint, and script/lint is a `docker build` of Dockerfile.lint,
+# so `RUN make check` here would be docker-in-docker inside a bare
+# alpine with no docker client and no daemon socket. The checks are
+# therefore split by where they run: the production build here, lint and
+# the prettier format check in Dockerfile.lint. Do not reintroduce a
+# `make check` (or a `make lint` / `make fmt-check`) line in this file.
+#
+# CI coverage is unaffected: script/cibuild builds this image and then
+# calls script/lint and script/fmt-check, so every check in `make check`
+# still runs on every push -- see script/cibuild.
 #
 # Build this only via script/cibuild or script/docker: both pass the
 # CHECK_EPOCH build argument that this file requires, and a bare
@@ -49,8 +51,5 @@ COPY . .
 ARG CHECK_EPOCH
 RUN [ -n "$CHECK_EPOCH" ] || exit 1
 
-# The individual non-lint checks - build fails if either fails. Invoked
-# as script/ entrypoints rather than `make check` for the reason in the
-# header comment above.
-RUN echo "check epoch: ${CHECK_EPOCH}" && script/test
-RUN script/fmt-check
+# Run the production build - build fails if the site does not build.
+RUN echo "check epoch: ${CHECK_EPOCH}" && make test

Dockerfile.lint

@@ -1,64 +1,85 @@
-# Lint-only image. `script/lint` builds this file and nothing else: the
-# lint runs as a build step, so a successful build IS a clean lint.
+# Lint image. Every lint-class check for this repo runs here and
+# nowhere else: the checks are build steps, so a successful build IS a
+# clean lint. There is no host lint path and no "am I already in a
+# container?" bypass -- script/lint and script/fmt-check are reduced to
+# building the stage below that carries their check.
 #
-# One stage, deliberately. A whole-file `docker build -f Dockerfile.lint .`
-# builds only the file's LAST stage, and sibling stages off a shared base
-# have no ordering edge between them, so a second stage sitting beside
-# this one would be silently skipped by exactly the invocation the
-# canonical org-wide `script/lint` uses -- a green that linted nothing,
-# which is the failure mode this file exists to prevent. With a single
-# stage there is nothing to skip and `script/lint` needs no `--target`.
-# If a second check is ever added here it must be chained (`FROM lint AS
-# ...`) or carry an explicit ordering edge, never left as a sibling.
+# Build this only via script/lint or script/fmt-check: both pass the
+# CHECK_EPOCH build argument that the stages here require, and a bare
+# `docker build -f Dockerfile.lint .` fails by design. See the guards.
 #
-# Only linting is containerised (owner ruling, 2026-08-10: "fmt and fmt
-# check arent docker, just linting"). script/fmt and script/fmt-check run
-# on the host, and the main Dockerfile runs the production build and the
-# format check directly -- see the comment there.
-#
-# The lint is invoked directly below rather than through `make lint` or
-# `script/lint`. That is not a style choice: `script/lint` IS this build,
-# so calling it from inside would recurse into a docker build with no
-# daemon.
-#
-# This repo's lint is a Hugo build that fails on any build error and on
-# any Hugo warning. --printPathWarnings makes Hugo warn when two pages
-# write the same output path, and --panicOnWarning turns that warning,
-# and every other one, into a failed build. Only the lint is this strict:
-# script/test and the deploy build stay plain `hugo --minify`.
+# Why this is a separate file from the main Dockerfile, and why that
+# one no longer runs `make check`: `make check` runs script/lint, and
+# script/lint is now a `docker build`. A `RUN make check` in an image
+# would therefore be docker-in-docker inside a bare alpine with no
+# docker client and no daemon socket. The checks are split by where
+# they run instead -- the main Dockerfile runs the production build,
+# this file runs lint and the format check -- so no image ever shells
+# back into `make check`. script/cibuild drives all of them, so CI
+# coverage is unchanged.
 
 # alpine 3.21, 2026-02-28
-FROM alpine@sha256:c3f8e73fdb79deaebaa2037150150191b9dcbfba68b4a46d70103204c53f4709
+FROM alpine@sha256:c3f8e73fdb79deaebaa2037150150191b9dcbfba68b4a46d70103204c53f4709 AS base
 
 WORKDIR /src
 
 # Keep these four instructions byte-identical to the main Dockerfile's,
 # in the same order: Docker keys layers on the instruction chain, not on
-# the file they live in, so an identical prefix means this build is a
+# the file they live in, so an identical prefix means this stage is a
 # cache hit against the main image's layers. script/bootstrap compiles
 # the pinned Hugo from source, which is by far the most expensive step
-# here, and it must not be paid twice. The dependency layer is also
-# deliberately above the ARG below, so it stays cached and only the lint
-# step re-runs on every invocation.
+# here, and it must not be paid twice.
 COPY script/ script/
 RUN script/bootstrap
 
 COPY . .
 
+# --- lint -------------------------------------------------------------
+#
+# The site's lint is a clean build that surfaces broken internal links
+# and template path problems: `hugo` fails on build errors and
+# --printPathWarnings reports render-target collisions.
+#
+# The flags are inlined rather than reached through `RUN script/lint`,
+# and that is forced, not lazy: script/lint is the docker build that
+# produces this stage, so calling it here is the recursion described
+# above. Keep these flags in sync with script/lint's documentation.
+FROM base AS lint
+
 # CHECK_EPOCH is a per-invocation nonce supplied by script/lint. Without
-# it an unchanged tree serves the lint layer from cache: the lint never
+# it an unchanged tree serves the check layer from cache: the lint never
 # executes and the build still exits 0, which is precisely the false
 # green this repo already fixed once in the main Dockerfile. Caching is
-# explicitly waived for lint, so the value is expanded into the linted
+# explicitly waived for lint, so the value is expanded into the checked
 # command as well as the guard -- two independent value-keyed
 # invalidation points, so a cache miss never depends on BuildKit's
-# treatment of an unreferenced ARG, and the epoch is visible in the build
-# log. Declared with no default: a default is a constant, and a constant
-# is a stable cache key. The guard makes a bare
-# `docker build -f Dockerfile.lint .` fail loudly instead of silently
-# reusing the empty (and therefore stable) cache key. Keep both
-# references.
+# treatment of an unreferenced ARG. Declared with no default: a default
+# is a constant, and a constant is a stable cache key. ARG is
+# stage-scoped, so the fmt-check stage below declares its own.
+#
+# Everything above this line still caches, so script/bootstrap is not
+# rebuilt.
 ARG CHECK_EPOCH
 RUN [ -n "$CHECK_EPOCH" ] || exit 1
-RUN echo "lint epoch: ${CHECK_EPOCH}" && \
-    hugo --minify --printPathWarnings --panicOnWarning
+RUN echo "lint epoch: ${CHECK_EPOCH}" && hugo --minify --printPathWarnings
+
+# --- fmt-check --------------------------------------------------------
+#
+# The read-only prettier check over this repo's markdown and CSS. Same
+# version, same scope and same flags as script/fmt, which is the
+# authoritative copy and stays on the host because it writes to the
+# working tree; keep the two in sync. The exclusions live in
+# .prettierignore with the reason for each.
+#
+# Built as a sibling of `lint` rather than stacked on top of it so that
+# a --target build runs exactly one check, and so a lint failure and a
+# formatting failure are reported independently.
+FROM base AS fmt-check
+
+# Same nonce, same reasoning as the lint stage above. Declared again
+# because ARG does not cross a FROM.
+ARG CHECK_EPOCH
+RUN [ -n "$CHECK_EPOCH" ] || exit 1
+RUN echo "fmt-check epoch: ${CHECK_EPOCH}" && \
+    npx --yes "prettier@3.4.2" --check \
+    '**/*.md' '**/*.css' --tab-width 4 --prose-wrap always

Makefile

@@ -1,4 +1,4 @@
-.PHONY: bootstrap setup test lint fmt fmt-check check docker cibuild precommit hooks serve
+.PHONY: bootstrap setup test lint fmt fmt-check check docker hooks serve
 
 bootstrap:
 	@script/bootstrap
@@ -24,12 +24,6 @@ check:
 docker:
 	@script/docker
 
-cibuild:
-	@script/cibuild
-
-precommit:
-	@script/precommit
-
 hooks:
 	@script/install-precommit
 

README.md

@@ -39,13 +39,12 @@ build, and the formatting check:
 make check
 ```
 
-The lint runs inside Docker, so `make check` needs a working Docker daemon;
-there is no host fallback. On a machine that has never built the image, the
-first run compiles the pinned Hugo from source, which takes minutes; later runs
-reuse that cached layer.
+The lint build and the formatting check run inside Docker, so `make check` needs
+a working Docker daemon; there is no host fallback.
 
 `make fmt` rewrites the repo's markdown and CSS to the project's prettier
-settings; run it if `make check` fails on formatting.
+settings; run it if `make check` fails on formatting. It runs on the host,
+because it writes to your working tree.
 
 To contribute to this site, contact **sneak@sneak.berlin** for git repository
 access.
@@ -66,52 +65,41 @@ provide:
   git pre-commit hook
 - `script/test` — the correctness check: a clean `hugo --minify` production
   build
-- `script/lint` — a Hugo build that fails on any Hugo warning, including two
-  pages that write the same output path, run inside Docker: it builds
-  `Dockerfile.lint`, where the lint is a build step, so a successful build is a
-  clean lint
+- `script/lint` — a clean build that surfaces broken links and path collisions,
+  run inside Docker: it builds the `lint` stage of `Dockerfile.lint`, where the
+  check is a build step, so a successful build is a clean lint
 - `script/fmt` — format every markdown and CSS file in the repo with prettier;
-  the exclusions live in `.prettierignore` with the reason for each
-- `script/fmt-check` — check that formatting (read-only)
+  the exclusions live in `.prettierignore` with the reason for each. The one
+  prettier entrypoint that runs on the host, because it writes to your working
+  tree
+- `script/fmt-check` — check that formatting (read-only), also inside Docker:
+  the `fmt-check` stage of `Dockerfile.lint`
 - `script/check` — run `script/test`, `script/lint`, then `script/fmt-check`;
   modifies no tracked files
 - `script/docker` — build the Docker image tagged with the project name
-- `script/cibuild` — the CI build: `script/lint` first, for fail-fast feedback,
-  then the main image, which runs the non-lint checks
+- `script/cibuild` — the CI build: the main image (the production build), then
+  `script/lint` and `script/fmt-check`
 - `script/install-precommit` — install the git pre-commit hook that runs
-  `script/precommit`
+  `script/check`
 
-Each of those has a Makefile shim of the same name — `make bootstrap`,
-`make setup`, `make test`, `make lint`, `make fmt`, `make fmt-check`,
-`make check`, `make docker`, `make cibuild` — with one exception:
-`script/install-precommit` is `make hooks`. `script/precommit`, which is what
-the installed hook runs, is `make precommit`. Prefer the make targets; the
-`Makefile` lists the operations you are expected to run.
+Every lint run for this repo happens inside a container. `script/lint` and
+`script/fmt-check` have no host path and no "already inside a container?"
+branch, so what a developer runs and what CI runs are the same build.
 
-`make cibuild` is the slowest target: it is the only one that runs two container
-builds, the lint image first and then the main image. The two share the
-`script/bootstrap` layer byte-for-byte, so the pinned-Hugo compile described
-above is paid once per machine rather than twice, and once that layer is cached
-a full `make cibuild` takes seconds. That is the cost of the CI build, not a
-sign of a problem.
+That is also why the main `Dockerfile` runs `make test` rather than
+`make check`: `make check` calls `script/lint`, which is itself a
+`docker build`, so a `make check` inside an image would be docker-in-docker in a
+bare Alpine with no docker client and no daemon socket. The checks are split by
+where they run — the production build in `Dockerfile`, lint and the format check
+in `Dockerfile.lint` — and `script/cibuild` drives all three, so CI coverage is
+unchanged.
 
-Every lint run for this repo happens inside a container, and only the lint does.
-`script/lint` has no host path and no "already inside a container?" branch, so
-what a developer runs and what CI runs are the same build. `script/fmt` and
-`script/fmt-check` run on the host: a formatting check is not a lint.
-
-That is also why the main `Dockerfile` runs `script/test` and `script/fmt-check`
-rather than `make check`. `make check` calls `script/lint`, which is itself a
-`docker build`, so a `make check` inside an image would attempt a docker build
-in a bare Alpine with no docker client and no daemon socket. The lint is not
-skipped — `script/cibuild` runs it first, in its own container, before the main
-image build starts.
-
-Build any image through `script/cibuild`, `script/docker` or `script/lint` only.
-All three pass a per-invocation `CHECK_EPOCH` build argument that the
-Dockerfiles require, so a check layer can never be served from cache — without
-it Docker returns a green it did not earn. A bare `docker build` fails closed on
-the `CHECK_EPOCH` guard rather than caching its way to a false success.
+Build any image through `script/cibuild`, `script/docker`, `script/lint` or
+`script/fmt-check` only. All four pass a per-invocation `CHECK_EPOCH` build
+argument that the Dockerfiles require, so a check layer can never be served from
+cache — without it Docker returns a green it did not earn. A bare `docker build`
+fails closed on the `CHECK_EPOCH` guard rather than caching its way to a false
+success.
 
 A convenience `make serve` target runs `hugo server` for local preview.
 

TODO.md

@@ -20,116 +20,62 @@ wrangler CLI install, an exact version), and the Hugo that builds the published
 site is a deliberate pinned version rather than whatever the base image's
 package repo serves. The site now ships a Cloudflare Pages `_headers` file, so
 its response security headers are declared in the repo instead of being whatever
-the edge defaults to, and confirmed live in production on both hostnames. The
-lint now runs inside a container and nowhere else: `script/lint` is a build of
-`Dockerfile.lint`, with no host path to fall back to. Formatting is not a lint
-and stays on the host. Every entrypoint the README documents now has a
-`Makefile` target, so the org-wide "use make targets, never the underlying tool"
-rule is satisfiable for the CI build as well as for the everyday checks.
+the edge defaults to — unverified in production until the next deploy. Every
+lint-class check now runs inside a container and nowhere else: `script/lint` and
+`script/fmt-check` build stages of `Dockerfile.lint`, with no host path to fall
+back to.
 
 # Next Step
 
-Make the prettier scope's exclusion of dot-directories explicit instead of
-leaning on `.gitignore` (https://git.eeqj.de/sneak/lora.vegas/issues/33).
+Add the missing `cibuild` and `precommit` shims to the `Makefile`, so that every
+documented entrypoint has a make target and the documented "always use make
+targets" rule is actually satisfiable
+(https://git.eeqj.de/sneak/lora.vegas/issues/34). Done when `make cibuild` and
+`make precommit` exist, are declared `.PHONY`, and the README Entrypoints
+section matches.
 
 # Completed Steps
 
-- 2026-11-05: added a `notify` job to the deploy workflow that posts an ntfy
-  message on every deploy, success or failure (closes
-  https://git.eeqj.de/sneak/lora.vegas/issues/48). The job runs `if: always()`
-  so it fires whether the build and deploy passed or failed, and the Notify step
-  carries `continue-on-error: true` so a notification that cannot be sent leaves
-  the deploy's result exactly as it was. On success the message body is the
-  short commit plus the site URL, and on failure it is a link to the run. The
-  topic comes from the `NTFY_TOPIC` Actions secret, which sneak must add before
-  anything is sent; the topic is never written in the repository. Both curl
-  calls carry `--max-time 10` so a stalled ntfy cannot hold the run open
-- 2026-10-21: made `script/lint` fail when two pages write the same output path
-  (https://git.eeqj.de/sneak/lora.vegas/issues/25). `Dockerfile.lint` now runs
-  `hugo --minify --printPathWarnings --panicOnWarning`. `--printPathWarnings`
-  makes Hugo warn about the collision, and `--panicOnWarning` turns that warning
-  and every other Hugo warning into a failed lint; until now the warning was
-  printed and the lint still exited 0. `script/test` and the deploy build stay
-  plain `hugo --minify`
-- 2026-08-10: added the `cibuild` and `precommit` targets to the `Makefile`
-  (https://git.eeqj.de/sneak/lora.vegas/issues/34). Both scripts already
-  existed, were already documented, and are the two entrypoints a contributor is
-  most likely to be told to run — the CI build and what the pre-commit hook runs
-  — yet neither had a make target, so the standing rule to use make targets
-  rather than the underlying tool could not be followed for either. It bit
-  reviewers twice, most recently on
-  https://git.eeqj.de/sneak/lora.vegas/pulls/32, and it bit hardest for the
-  build: since https://git.eeqj.de/sneak/lora.vegas/issues/30 a bare
-  `docker build .` fails closed on the `CHECK_EPOCH` guard, so `script/cibuild`
-  is one of only three supported ways to build an image and was the only one
-  without a target. The two targets are thin shims in the existing style and
-  change nothing about what the scripts do. `.PHONY` was already complete and
-  now lists both. `README.md`'s Entrypoints section gained the script-to-target
-  mapping, including the two names that do not match —
-  `script/install-precommit` is `make hooks`, and `script/precommit` is
-  `make precommit` — plus a note that `make cibuild` is the slowest target
-  because it is the only one that runs two container builds, while still taking
-  seconds once the shared `script/bootstrap` layer is cached. The section's
-  pre-existing `script/install-precommit` bullet, which claimed the installed
-  hook runs `script/check`, now says `script/precommit`, which is what the
-  script actually writes into `.git/hooks/pre-commit`. The stale Future Step
-  asking for post-deploy confirmation of `static/_headers` is dropped here: it
-  was confirmed live on both hostnames
-  (https://git.eeqj.de/sneak/lora.vegas/issues/14). Verified that `make cibuild`
-  earns its green rather than replaying a warm cache: one invocation ran both
-  builds for real in sequence, each with its own distinct `CHECK_EPOCH`, with
-  `RUN script/bootstrap` `CACHED` above and no check layer cached below it — the
-  `Dockerfile.lint` build echoed its epoch and printed hugo's own build table,
-  then the main image build echoed a different epoch, ran `script/test` for the
-  production build and `script/fmt-check` for the formatting check.
-  `make precommit` passes on a clean tree, and `make check` passes
-- 2026-08-10: moved the lint into Docker
+- 2026-08-10: moved every lint-class check into Docker
   (https://git.eeqj.de/sneak/lora.vegas/issues/38). A new root `Dockerfile.lint`
-  runs `hugo --minify --printPathWarnings` as a build step, so a successful
-  build is a clean lint, and `script/lint` is nothing but a build of that file —
-  no host path and deliberately no "already inside a container?" branch, which
-  would be a host lint path in disguise. The containerisation boundary is lint
-  only, per the owner ruling of the same day: formatting is not a lint, so
-  `script/fmt` and `script/fmt-check` stay on the host. `Dockerfile.lint` has
-  exactly one stage on purpose. A whole-file `docker build -f Dockerfile.lint .`
-  builds only the file's last stage, and sibling stages off a shared base have
-  no ordering edge, so any second stage beside the lint would be silently
-  skipped by the invocation the canonical org-wide `script/lint` uses — a green
-  that linted nothing. With one stage there is nothing to skip and `script/lint`
-  needs no `--target`. Its first four instructions are byte-identical to the
-  main `Dockerfile`'s, so the expensive `RUN script/bootstrap` layer that
-  compiles Hugo from source is shared between the two images rather than paid
-  twice. The recursion this creates was resolved by direction, not detection:
-  `make check` calls `script/lint`, so the main `Dockerfile` can no longer
-  `RUN make check` — that would attempt a docker build inside a build step, in a
-  bare Alpine with no docker client and no daemon socket. It runs the individual
-  non-lint checks instead, `script/test` and `script/fmt-check`, matching the
-  canonical shape upstream, and `script/cibuild` runs `script/lint` first for
-  fail-fast feedback before the main image build starts. So CI still covers all
-  three checks and cannot drift from what a developer runs. Caching is waived
-  for the lint exactly as the main `Dockerfile` already does it:
-  `ARG CHECK_EPOCH` with no default, guarded with
-  `[ -n "$CHECK_EPOCH" ] || exit 1`, and the value expanded into the linted
-  command as well as the guard, so invalidation never rests on BuildKit's
-  handling of an unreferenced `ARG`. Every image-building entrypoint generates
-  and passes it — `script/cibuild`, `script/docker`, `script/lint` — which is
-  the failure mode this repo already hit once, a Dockerfile guard asserting a
-  property one entrypoint did not supply. `script/lint` builds with
-  `--output type=cacheonly`: the build is run for its exit status, not for an
-  image, and since the lint layer is cache-busted every run an exporting build
-  would leave one dangling image per lint on a host shared with other work.
-  Verified rather than assumed: two consecutive `script/lint` runs on an
-  unchanged tree both executed hugo for real (second run 0.85s wall,
-  `RUN script/bootstrap` `CACHED`, distinct epochs echoed, real build tables
-  printed); a whole-file `docker build -f Dockerfile.lint .` with the argument
-  and no `--target` ran the lint for real, which is the regression test for the
-  skipped-sibling hazard; a bare build with no argument failed closed on the
-  guard; a planted template error failed the lint with hugo's own render error
-  and made `script/cibuild` exit in 0.6s without the main image build starting
-  at all; a planted over-long line failed the host `script/fmt-check` with
+  carries the checks as build steps — a `lint` stage running
+  `hugo --minify --printPathWarnings` and a `fmt-check` stage running the
+  prettier check — on a shared `base` stage whose first four instructions are
+  byte-identical to the main `Dockerfile`'s, so the expensive
+  `RUN script/bootstrap` layer that compiles Hugo from source is a cache hit
+  against the main image rather than a second build of the same thing.
+  `script/lint` and `script/fmt-check` are now nothing but a `docker build` of
+  their stage; there is no host path and deliberately no "already inside a
+  container?" branch, which would be a host lint path in disguise. The recursion
+  this creates was resolved by splitting the checks by where they run rather
+  than by adding an escape hatch: `make check` runs `script/lint`, so the main
+  `Dockerfile` can no longer `RUN make check` — that would be docker-in-docker
+  inside a bare Alpine with no docker client and no daemon socket, and
+  `script/cibuild` is what CI runs on every push. The main `Dockerfile`
+  therefore runs `make test`, the production build, and `script/cibuild` builds
+  it and then calls `script/lint` and `script/fmt-check`, so CI still covers all
+  three and cannot drift from what a developer runs. `script/fmt` stays on the
+  host because it rewrites the working tree, which makes it the authoritative
+  copy of the prettier version and flags that the `fmt-check` stage duplicates;
+  both sides carry a keep-in-sync note, and that duplication is forced, since
+  any `RUN script/fmt-check` inside the image is the recursion again. Caching is
+  waived for the checks exactly as the main `Dockerfile` already does it:
+  `ARG CHECK_EPOCH` with no default, declared and guarded separately in each
+  stage because `ARG` does not cross a `FROM`, with the value expanded into the
+  checked command as well as the guard. All four image-building entrypoints now
+  generate and pass it — `script/cibuild`, `script/docker`, `script/lint`,
+  `script/fmt-check` — which is the failure mode this repo already hit once, a
+  Dockerfile guard asserting a property one entrypoint did not supply. Verified
+  rather than assumed: two consecutive `script/lint` runs on an unchanged tree
+  both executed hugo for real (second run 2.9s wall, `RUN script/bootstrap`
+  `CACHED`, distinct epoch echoed, `Total in 37 ms` printed), a constant-epoch
+  counterfactual restored the false green (exit 0 in 0.25s, lint layer `CACHED`,
+  no hugo output at all), an empty epoch failed closed on the guard, a broken
+  template failed the lint stage with hugo's own render error, and an over-long
+  line appended to `README.md` failed the fmt-check stage with
   `[warn] README.md`; both violations were reverted and re-run clean. Not
-  changed here, and still true: the lint fails on hugo build errors but not on
-  render-target collisions, which `--printPathWarnings` only prints
+  changed here, and still true: the lint stage fails on hugo build errors but
+  not on render-target collisions, which `--printPathWarnings` only prints
   (https://git.eeqj.de/sneak/lora.vegas/issues/25) — containerising the run
   neither fixes nor worsens that
 - 2026-08-10: added the `LICENSE` file and made the README say what it says
@@ -330,6 +276,8 @@ leaning on `.gitignore` (https://git.eeqj.de/sneak/lora.vegas/issues/33).
 Startable work first. Everything under "Blocked" waits on somebody or something
 outside this repo, so nothing there may be picked up as the Next Step.
 
+- Make the prettier scope's exclusion of dot-directories explicit instead of
+  leaning on `.gitignore` (https://git.eeqj.de/sneak/lora.vegas/issues/33)
 - Fix the README's SSH-only clone URL, and add the two entrypoints the
   Entrypoints section omits, `script/precommit` and `script/projectname`
   (https://git.eeqj.de/sneak/lora.vegas/issues/36)
@@ -345,6 +293,10 @@ outside this repo, so nothing there may be picked up as the Next Step.
 
 ## Blocked
 
+- Decide whether `script/lint` should fail on render-target collisions rather
+  than only print them; `--printPathWarnings` exits 0 today, so the signal is
+  reported and not enforced. Owner call, since it changes what the gate rejects
+  (https://git.eeqj.de/sneak/lora.vegas/issues/25)
 - Move the artifact actions in `.gitea/workflows/deploy.yml` to v4 once this
   Gitea Actions instance serves the v4 artifact protocol; they are pinned on the
   deprecated v3 line because v4 fails here
@@ -355,6 +307,17 @@ outside this repo, so nothing there may be picked up as the Next Step.
 - Delete the stale remote branches `feat/initial-site` and `security-audit`;
   only the owner can remove them
   (https://git.eeqj.de/sneak/lora.vegas/issues/15)
+- After the next deploy, confirm the `_headers` file actually took effect, on
+  both `https://lora.vegas/` and `https://www.lora.vegas/`: `curl -sSI` against
+  each must show `strict-transport-security` or `content-security-policy`.
+  Cloudflare Pages silently ignores a malformed `_headers`, and checking
+  `x-content-type-options` would pass either way because the edge sends it
+  regardless. `www` has to be checked too and not just the apex: dropping
+  `includeSubDomains` rests on `www.lora.vegas` being served by this same Pages
+  project, which was established behaviourally from identical response bodies
+  rather than from the Cloudflare dashboard. If `www` turns out not to be
+  covered, the `includeSubDomains` decision has to be revisited
+  (https://git.eeqj.de/sneak/lora.vegas/issues/14)
 - Decide the HSTS `includeSubDomains` and `preload` posture for `lora.vegas`.
   Both are owner calls: neither can be walked back inside the max-age window,
   and `includeSubDomains` binds hostnames this repo does not control

script/check

@@ -2,15 +2,12 @@
 # script/check: run all checks. Our own extension to
 # scripts-to-rule-them-all. Must not modify any tracked files. Runs the
 # canonical order: the clean production build, then the lint build that
-# fails on any Hugo warning, then the read-only formatting check.
+# reports path warnings, then the read-only formatting check.
 #
-# The lint - and only the lint - runs inside Docker: it is a build of
-# Dockerfile.lint, so this script needs a working docker daemon and has
-# no host fallback to drop back to. Budget for the cold case: the first
-# lint on a machine with no cached script/bootstrap layer compiles the
-# pinned Hugo from source, which takes minutes. That cost falls on the
-# pre-commit hook too, since it runs this script. Every later run reuses
-# that layer and only the lint step re-executes.
+# The last two run inside Docker (they build stages of Dockerfile.lint),
+# so this script needs a working docker daemon. That is deliberate:
+# every lint run for this repo happens in a container, and there is no
+# host fallback to drop back to.
 set -eu
 
 SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd -P)"

script/cibuild

@@ -1,30 +1,28 @@
 #!/bin/sh
 # script/cibuild: run the CI build. The Gitea workflow runs this on
-# push, and it is the single entrypoint that covers everything. Two
-# container builds, in order:
+# push, and it is the single entrypoint that covers everything:
 #
-#   1. script/lint, which builds Dockerfile.lint -- the lint runs as a
-#      build step there
-#   2. the main Dockerfile, which runs the non-lint checks: the clean
-#      `hugo --minify` production build (script/test) and the read-only
-#      prettier check (script/fmt-check)
+#   1. the main Dockerfile, which runs the clean `hugo --minify`
+#      production build (`make test`)
+#   2. script/lint, which builds Dockerfile.lint's `lint` stage
+#   3. script/fmt-check, which builds Dockerfile.lint's `fmt-check`
+#      stage
 #
-# Lint goes first, for fail-fast feedback: on a runner with no cached
-# script/bootstrap layer the main image compiles Hugo from source, and a
-# lint failure should not wait behind that. It is a separate build
-# rather than a step inside the main image because script/lint is itself
-# a `docker build`, and a docker build cannot run a docker build. See
-# Dockerfile.lint for the full reasoning.
+# Steps 2 and 3 are delegated to the same scripts a developer runs, so
+# CI cannot drift from `make check`. They are separate builds rather
+# than a `RUN make check` inside the main image because script/lint is
+# itself a `docker build`: shelling back into `make check` from an image
+# would be docker-in-docker inside a bare alpine with no docker client
+# and no daemon socket. See Dockerfile.lint for the full reasoning.
 #
-# The lint is delegated to the same script a developer runs, so CI
-# cannot drift from `make check`.
-#
-# Neither build implies a passing check without CHECK_EPOCH. Docker keys
-# the check layers on content, so on an unchanged tree they are served
-# from cache: nothing executes and the build still exits 0. Passing a
-# value that differs on every invocation invalidates those layers and
-# everything below them, while the script/bootstrap toolchain layer
-# above keeps caching. script/lint does the same for its own build.
+# The main image build below only implies a passing production build
+# because of CHECK_EPOCH. Docker keys the `RUN make test` layer on
+# content, so on an unchanged tree it is served from cache: the build
+# never executes and the image build still exits 0. Passing a value that
+# differs on every invocation invalidates that layer and everything
+# below it, while the script/bootstrap toolchain layer above it keeps
+# caching. script/lint and script/fmt-check each do the same for their
+# own stage.
 set -eu
 
 SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd -P)"
@@ -32,8 +30,6 @@ ROOT="$(cd "$SCRIPT_DIR/.." && pwd -P)"
 
 main() {
     cd "$ROOT"
-    "$SCRIPT_DIR/lint"
-
     # Assigned to a variable rather than substituted inline in the
     # argument list: a command substitution that fails inside an
     # argument does not trip `set -e`, so the inline form would quietly
@@ -44,6 +40,9 @@ main() {
     # 0, so `$$` is appended to cover that degradation.
     epoch="$(date +%s%N)$$"
     docker build --build-arg CHECK_EPOCH="$epoch" .
+
+    "$SCRIPT_DIR/lint"
+    "$SCRIPT_DIR/fmt-check"
 }
 
 main "$@"

script/fmt

@@ -7,10 +7,11 @@
 # Go templates and not HTML, and content/, whose reformatting was
 # measured to change the rendered page.
 #
-# Both prettier entrypoints run on the host: only linting is
-# containerised (owner ruling, 2026-08-10), and formatting is not a
-# lint. Keep the version, scope and flags here in sync with
-# script/fmt-check.
+# This runs on the host, unlike the read-only check: it rewrites the
+# working tree, which a container build cannot do. It is therefore the
+# authoritative copy of the prettier version, scope and flags -- the
+# fmt-check stage of Dockerfile.lint duplicates them and must be kept in
+# sync with this file.
 set -eu
 
 ROOT="$(cd "$(dirname "$0")/.." && pwd -P)"

script/fmt-check

@@ -1,22 +1,27 @@
 #!/bin/sh
 # script/fmt-check: check the formatting of this repo's markdown and
-# CSS (read-only). Same scope and same settings as script/fmt - keep
-# the two in sync - but fails instead of writing.
+# CSS (read-only). Like script/lint, it runs only in Docker: it builds
+# the `fmt-check` stage of Dockerfile.lint, where prettier runs as a
+# build step. Leaving prettier to run on the host here would have left
+# `make check` with a host lint path, which is the thing being removed.
 #
-# This runs on the host, not in a container: only linting is
-# containerised (owner ruling, 2026-08-10), and a formatting check is
-# not a lint. Running here also keeps it usable offline, since npx
-# reuses ~/.npm/_npx after the first run.
+# The version, scope and flags live in Dockerfile.lint and must stay in
+# sync with script/fmt, which is the authoritative copy and stays on the
+# host because it writes to the working tree.
+#
+# Dockerfile.lint requires the CHECK_EPOCH build argument, generated
+# here exactly as script/cibuild generates it -- see that script for why
+# the check layer must not be allowed to cache, and why the value is
+# built in an assignment rather than inline.
 set -eu
 
 ROOT="$(cd "$(dirname "$0")/.." && pwd -P)"
 
-PRETTIER_VERSION="3.4.2"
-
 main() {
     cd "$ROOT"
-    npx --yes "prettier@${PRETTIER_VERSION}" --check \
-        '**/*.md' '**/*.css' --tab-width 4 --prose-wrap always
+    epoch="$(date +%s%N)$$"
+    docker build -f Dockerfile.lint --target fmt-check \
+        --build-arg CHECK_EPOCH="$epoch" .
 }
 
 main "$@"

script/lint

@@ -1,21 +1,16 @@
 #!/bin/sh
 # script/lint: run the lint. This Hugo site has no dedicated linter, so
-# the lint is a Hugo build that fails on any build error and on any Hugo
-# warning, including two pages that write the same output path. Where
-# it runs is not negotiable. Every lint run happens inside a Docker
-# container, so this script does nothing except build Dockerfile.lint.
-# The lint is a build step there, so a successful build is a clean lint.
-# There is deliberately no host fallback and no "already inside a
-# container?" branch: either would be a host lint path wearing a
-# disguise.
-#
-# No --target: Dockerfile.lint has exactly one stage, so the whole-file
-# build IS the lint. See that file for why a second, sibling stage would
-# be a silent skip.
+# the lint gate is a clean build that surfaces broken internal links and
+# template path problems -- but where it runs is not negotiable: every
+# lint run happens inside a Docker container, so this script does
+# nothing except build the `lint` stage of Dockerfile.lint. The check is
+# a build step there, so a successful build is a clean lint. There is
+# deliberately no host fallback and no "already inside a container?"
+# branch: either would be a host lint path wearing a disguise.
 #
 # Dockerfile.lint requires the CHECK_EPOCH build argument, generated
 # here exactly as script/cibuild generates it -- see that script for why
-# the lint layer must not be allowed to cache, and why the value is
+# the check layer must not be allowed to cache, and why the value is
 # built in an assignment rather than inline.
 set -eu
 
@@ -24,18 +19,8 @@ ROOT="$(cd "$(dirname "$0")/.." && pwd -P)"
 main() {
     cd "$ROOT"
     epoch="$(date +%s%N)$$"
-    # --output type=cacheonly: this build is run for its exit status,
-    # not for an image. Because the lint layer is cache-busted on every
-    # invocation the result is a new image every time, and an untagged
-    # build would leave one dangling image per lint run on a host shared
-    # with other work. cacheonly keeps the build cache (so
-    # script/bootstrap still hits) and exports nothing. Failures still
-    # propagate: an empty CHECK_EPOCH or a failing lint exits non-zero.
-    docker build \
-        --build-arg CHECK_EPOCH="$epoch" \
-        --output type=cacheonly \
-        -f Dockerfile.lint \
-        .
+    docker build -f Dockerfile.lint --target lint \
+        --build-arg CHECK_EPOCH="$epoch" .
 }
 
 main "$@"