diff --git a/.gitea/workflows/deploy-services.yml b/.gitea/workflows/deploy-services.yml
index bdecf2e7..55ca72a2 100644
--- a/.gitea/workflows/deploy-services.yml
+++ b/.gitea/workflows/deploy-services.yml
@@ -1,17 +1,23 @@
-# Deploy-only: bring up the two unom-1 pieces that live in THIS repo but whose normal
-# deploys are coupled to heavy build workflows — docs to docker.yml's 5-image matrix,
-# the flatpak server to flatpak.yml's full flatpak-builder run. This workflow does
-# NEITHER build: it just (re)places the compose files and pulls the already-published
-# images, so unom/infra's deploy-all can bring a fresh unom-1 fully up in a single
-# dispatch without triggering those rebuilds.
+# Deploy-only: bring up the unom-1 pieces that live in THIS repo but whose normal deploys
+# are coupled to heavy build workflows — docs to docker.yml's 5-image matrix, the flatpak
+# server to flatpak.yml's full flatpak-builder run, the nix cache to nix.yml's full Rust
+# build. This workflow does NONE of those builds: it just (re)places the compose files and
+# pulls the already-published images, so unom/infra's deploy-all can bring a fresh unom-1
+# fully up in a single dispatch without triggering those rebuilds.
 #
-#   docs     -> pulls git.unom.io/unom/punktfunk-docs:latest (built by docker.yml) and
-#               brings it up on :3220.
-#   flatpak  -> brings up the caddy:2-alpine static server on :3230. The OSTree repo
-#               CONTENT (./site) is NOT shipped here — it is regenerated by flatpak.yml
-#               on the next client build, or restored from the unom-1 backup
-#               (unom/infra scripts/restore-unom-1.sh, `files` tag). A fresh box serves
-#               an empty repo until then; that is expected.
+#   docs      -> pulls git.unom.io/unom/punktfunk-docs:latest (built by docker.yml) and
+#                brings it up on :3220.
+#   flatpak   -> brings up the caddy:2-alpine static server on :3230. The OSTree repo
+#                CONTENT (./site) is NOT shipped here — it is regenerated by flatpak.yml
+#                on the next client build, or restored from the unom-1 backup
+#                (unom/infra scripts/restore-unom-1.sh, `files` tag). A fresh box serves
+#                an empty repo until then; that is expected.
+#   nix-cache -> brings up the caddy:2-alpine Nix binary cache on :3250. Same content/config
+#                split: the cache CONTENT is republished by nix.yml on the next main push
+#                that moves the flake. An empty cache is harmless — every path 404s and
+#                users build from source, which is the pre-cache status quo.
+#   winget    -> brings up the winget REST source on :3240; catalogue shipped by
+#                windows-host.yml on stable tags.
 #
 # Dispatched by unom/infra scripts/deploy-all.sh: `dispatch-and-wait.sh punktfunk
 # deploy-services.yml`. Uses the same secret set docker.yml/flatpak.yml already rely on:
@@ -100,6 +106,46 @@ jobs:
             cd ~/unom-flatpak
             docker compose -f compose.production.yml up -d
 
+  nix-cache:
+    runs-on: ubuntu-24.04
+    timeout-minutes: 10
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Sync nix cache compose + server
+        uses: appleboy/scp-action@917f8b81dfc1ccd331fef9e2d61bdc6c8be94634 # v0.1.7
+        with:
+          host: ${{ inputs.deploy_host || secrets.DEPLOY_HOST }}
+          username: ${{ secrets.DEPLOY_USER }}
+          port: ${{ secrets.DEPLOY_PORT }}
+          key: ${{ secrets.DEPLOY_SSH_KEY }}
+          # Land all three flat in ~/unom-nix-cache/ (drop the packaging/nix/server/ prefix).
+          source: "packaging/nix/server/compose.production.yml,packaging/nix/server/Caddyfile,packaging/nix/server/prune.sh"
+          target: "~/unom-nix-cache"
+          strip_components: 3
+          overwrite: true
+
+      - name: Start nix binary cache server
+        uses: appleboy/ssh-action@0ff4204d59e8e51228ff73bce53f80d53301dee2 # v1.2.5
+        with:
+          host: ${{ inputs.deploy_host || secrets.DEPLOY_HOST }}
+          username: ${{ secrets.DEPLOY_USER }}
+          port: ${{ secrets.DEPLOY_PORT }}
+          key: ${{ secrets.DEPLOY_SSH_KEY }}
+          script: |
+            set -euo pipefail
+            # ./site (the cache CONTENT) is NOT shipped here — nix.yml rsyncs it on each main push
+            # that moves the flake, same content/config split as the flatpak repo and the winget
+            # catalogue. Ensure the bind-mount source exists so caddy starts; an empty cache is
+            # harmless, it just 404s every path and users build from source as they do today.
+            mkdir -p ~/unom-nix-cache/site/nar
+            cd ~/unom-nix-cache
+            docker compose -f compose.production.yml up -d
+            # A cache that 404s a miss is healthy; one that cannot answer at all is not.
+            sleep 3
+            curl -fsS http://127.0.0.1:3250/nix-cache-info \
+              || echo "NOTE: no cache content yet - push to main with the flake touched to populate it"
+
   winget:
     runs-on: ubuntu-24.04
     timeout-minutes: 10
diff --git a/.gitea/workflows/nix.yml b/.gitea/workflows/nix.yml
index d8a99526..310fbd18 100644
--- a/.gitea/workflows/nix.yml
+++ b/.gitea/workflows/nix.yml
@@ -4,8 +4,9 @@
 # `nix build .#punktfunk-web` was broken for 553 commits before anyone noticed (see the bun-nix job
 # in ci.yml for that story).
 #
-# Two tiers, because a full `nix flake check` builds the whole Rust workspace with crane and would
-# run for an hour on every push:
+# Three tiers, because a full `nix flake check` builds the whole Rust workspace with crane and would
+# run for an hour on every push — so the two cheap tiers gate every PR and the expensive one runs
+# only where its cost buys something (a published cache):
 #
 #   * eval  — `nix flake check --no-build`: instantiates every package, app, check and devShell
 #             without building them. Catches the failures that actually happen to this flake — a
@@ -32,15 +33,28 @@
 #             covers what the ci.yml drift gate cannot, e.g. a tarball the registry no longer
 #             serves, or the codegen going quietly message-less (see packages.nix's inlang note).
 #
-# The Rust packages (punktfunk-host, punktfunk-client) and punktfunk-gamescope are NOT built here.
-# They are the expensive ones and their inputs are already gated by the `rust` job in ci.yml; build
-# them by hand on a Nix box, or with the `build-rust` dispatch input below.
+#   * cache — PUSH TO MAIN ONLY. Builds the Rust packages + gamescope for real and publishes every
+#             punktfunk store path to the binary cache at https://nix.unom.io, so a NixOS user gets
+#             prebuilt binaries instead of an hour of rustc and a gamescope compile. This is the
+#             expensive tier and it is why the job timeout is 180 rather than 90.
 #
-# ⚠ punktfunk-gamescope deserves the dispatch run more than it looks: `host.gamescopeHdr` DEFAULTS
-# TRUE, so it is on the critical path of every `services.punktfunk.host.enable = true` build, while
-# being the one package nothing here compiles. It patches whatever gamescope the pinned nixpkgs
-# carries, so a nixpkgs bump — not a change of ours — is what breaks it, and the first person to
-# find out would be an operator whose system rebuild fails. Run the dispatch after a flake.lock bump.
+#             It needs NO extra trigger for releases: a release bumps the workspace version in
+#             Cargo.toml, which is already in the path filter below, so the tag's content is
+#             published by the version-bump commit on main.
+#
+#             Only OUR paths are published — see the step for why that is both correct and the
+#             difference between ~300 MB and several GB per publish.
+#
+# The Rust packages and punktfunk-gamescope are still not built on PRs: they are the expensive ones
+# and their inputs are already gated by the `rust` job in ci.yml. Build them on a PR by hand on a
+# Nix box, or with the `build-rust` / `build-gamescope` dispatch inputs below.
+#
+# ⚠ punktfunk-gamescope matters more than it looks: `host.gamescopeHdr` DEFAULTS TRUE, so it is on
+# the critical path of every `services.punktfunk.host.enable = true` build. It patches whatever
+# gamescope the pinned nixpkgs carries, so a nixpkgs bump — not a change of ours — is what breaks
+# it, and the first person to find out would be an operator whose system rebuild fails. The `cache`
+# tier now compiles it on every main push, so a flake.lock bump that breaks it goes red HERE; the
+# dispatch input below is for checking it on a branch before merging.
 #
 # ⚠ pull_request is deliberately present. flatpak.yml shipped with push-only triggers and manifest
 # breakage reached main invisibly for weeks — do not "simplify" this workflow by dropping it.
@@ -107,8 +121,15 @@ jobs:
       # real node (so actions/checkout works with no pre-checkout install dance), and audit.yml
       # already pulls it on this fleet, so it is proven to resolve here. Nix is installed below.
       image: node:22-bookworm
-    timeout-minutes: 90
+    # 180, not 90: the `cache` tier on a main push compiles the whole Rust workspace AND gamescope
+    # from source inside the nix sandbox, where the sccache every other Rust job leans on cannot
+    # reach (no network in a derivation, and RUSTC_WRAPPER is not set inside one).
+    timeout-minutes: 180
     env:
+      # Where the published cache lives on unom-1, and the URL users substitute from. Kept next to
+      # the flatpak repo (3230) and winget source (3240) — see packaging/nix/server/.
+      DEPLOY_DIR: unom-nix-cache
+      CACHE_URL: https://nix.unom.io
       # The flake needs both experimental features. Also baked into the installer's --extra-conf
       # below; this covers any step that shells out before that config is read.
       NIX_CONFIG: "experimental-features = nix-command flakes"
@@ -126,11 +147,12 @@ jobs:
     steps:
       - uses: actions/checkout@v4
 
-      # The Determinate installer needs curl + xz; git so nix can read the flake from the checkout.
-      # (node:22-bookworm is the full image and already has all three — this is belt-and-braces
-      # against a future slim-image swap, and costs one cached apt call.)
+      # The Determinate installer needs curl + xz; git so nix can read the flake from the checkout;
+      # rsync + ssh to ship the built cache to unom-1. (node:22-bookworm is the full image and
+      # already has all but rsync — this is belt-and-braces against a future slim-image swap, and
+      # costs one cached apt call.)
       - name: Installer prerequisites
-        run: apt-get update && apt-get install -y --no-install-recommends ca-certificates curl xz-utils git
+        run: apt-get update && apt-get install -y --no-install-recommends ca-certificates curl xz-utils git rsync openssh-client
 
       # `--init none` is the container mode: no systemd, no daemon. Running as root, nix then talks
       # to the store directly. Determinate Nix is also what the Nix box (.21) runs, so CI and the
@@ -183,6 +205,133 @@ jobs:
             || { echo "installed console is not a bun bundle" >&2; exit 1; }
           echo "bun packages OK: $web  $scripting"
 
+      # ── binary cache (push to main only) ───────────────────────────────────────────────────────
+      #
+      # Decided against a bucket on storage.unom.io even though sccache already uses it and the
+      # credentials already exist: it is local RustFS on the home uplink with no CDN in front, so
+      # every NixOS user's download would come off the same pipe every CI runner uses — and S3
+      # answers 403, not 404, for a missing key, which nix treats as a hard error rather than a
+      # cache miss (see packaging/nix/server/Caddyfile). unom-1 already serves the flatpak repo
+      # this way from a cloud IP; a Nix cache is the same static-files-behind-caddy shape.
+      #
+      # Gitea itself cannot host this at all: it has 23 package registry types and none is Nix, and
+      # the binary cache protocol wants fixed anonymous paths at a URL root (/nix-cache-info,
+      # /<hash>.narinfo, /nar/<hash>.nar.xz) that /api/packages/{owner}/generic/… cannot express.
+      - name: Cache publish preflight
+        id: cachecfg
+        if: ${{ github.event_name == 'push' }}
+        env:
+          NIX_CACHE_SIGNING_KEY: ${{ secrets.NIX_CACHE_SIGNING_KEY }}
+          DEPLOY_HOST: ${{ secrets.DEPLOY_HOST }}
+        # Guard BEFORE the build, not before the upload: an unconfigured cache must not cost an
+        # hour of rustc first. No-ops cleanly until the secret exists, exactly as flatpak.yml's
+        # repo deploy does, so this workflow stays green through setup.
+        run: |
+          set -eu
+          if [ -n "${NIX_CACHE_SIGNING_KEY:-}" ] && [ -n "${DEPLOY_HOST:-}" ]; then
+            echo "go=true" >> "$GITHUB_OUTPUT"
+          else
+            echo "go=false" >> "$GITHUB_OUTPUT"
+            echo "::warning::NIX_CACHE_SIGNING_KEY/DEPLOY_HOST not set — skipping the binary cache publish (see packaging/nix/README.md)."
+          fi
+
+      - name: Build the publishable packages
+        if: ${{ steps.cachecfg.outputs.go == 'true' }}
+        # Everything a user can install. punktfunk-gamescope earns its place here more than any
+        # other: host.gamescopeHdr DEFAULTS TRUE, so without it in the cache every
+        # `services.punktfunk.host.enable = true` still compiles a compositor from source.
+        run: |
+          "$NIX" build --print-build-logs \
+            .#punktfunk-host .#punktfunk-client .#punktfunk-tray \
+            .#punktfunk-web .#punktfunk-scripting .#punktfunk-gamescope
+          # This is now the heaviest job on the fleet — a full workspace build plus gamescope fills
+          # the store with tens of GB, and this fleet ran a runner out of disk on 2026-08-06. Record
+          # the headroom AFTER the build too, or a future "no space left on device" is a guess.
+          df -h / /nix /tmp || true
+
+      - name: Sign + publish to nix.unom.io
+        if: ${{ steps.cachecfg.outputs.go == 'true' }}
+        env:
+          NIX_CACHE_SIGNING_KEY: ${{ secrets.NIX_CACHE_SIGNING_KEY }}
+          DEPLOY_HOST: ${{ secrets.DEPLOY_HOST }}
+          DEPLOY_USER: ${{ secrets.DEPLOY_USER }}
+          DEPLOY_PORT: ${{ secrets.DEPLOY_PORT }}
+          DEPLOY_SSH_KEY: ${{ secrets.DEPLOY_SSH_KEY }}
+        run: |
+          # `set -eu`, NOT `set -euo pipefail`: act_runner may execute a step's `run:` under dash in
+          # these containers (see scripts/ci/ensure-sccache.sh), and dash dies on `-o pipefail` with
+          # "Illegal option". The two places below where a pipeline's LEFT side must be able to fail
+          # the step are written as redirects instead, so nothing depends on pipefail.
+          set -eu
+          PKGS=".#punktfunk-host .#punktfunk-client .#punktfunk-tray .#punktfunk-web .#punktfunk-scripting .#punktfunk-gamescope"
+
+          # 1) Pick what to publish. PUBLISH ONLY OUR OWN PATHS — this is the difference between
+          #    ~300 MB and several GB per run, and it is not a corner cut: a runtime closure here
+          #    is our binaries plus stock nixpkgs (ffmpeg, gtk4, glibc, …), and every stock path is
+          #    already on cache.nixos.org, served by a real CDN. Mirroring them onto unom-1 would
+          #    cost disk and home-to-cloud bandwidth to serve a WORSE copy of what users already
+          #    have. Nothing in nixpkgs is named punktfunk, so the name filter is exact.
+          paths="$("$NIX" path-info -r $PKGS | grep -- '-punktfunk' || true)"
+          [ -n "$paths" ] || { echo "::error::no punktfunk store paths in the closure — the name filter is broken"; exit 1; }
+          echo "$paths"
+          # The filter is a string match, so it would fail SILENTLY if a pname ever changed — and
+          # the package most likely to drift is gamescope, the most expensive one to lose. Assert
+          # every built output is actually covered rather than discovering it as a user rebuild.
+          for out in $("$NIX" build --print-out-paths $PKGS); do
+            printf '%s\n' "$paths" | grep -qxF "$out" \
+              || { echo "::error::$out is not matched by the '-punktfunk' filter — publish would silently omit it"; exit 1; }
+          done
+
+          # 2) Sign into a local binary cache. The secret is the whole `name:base64` line from
+          #    `nix key generate-secret`; the matching public key is what users pin (README).
+          KEYDIR="$(mktemp -d)"; chmod 700 "$KEYDIR"
+          printf '%s' "$NIX_CACHE_SIGNING_KEY" > "$KEYDIR/key"; chmod 600 "$KEYDIR/key"
+          printf '%s\n' "$paths" | xargs "$NIX" copy --to "file://$PWD/nix-cache?secret-key=$KEYDIR/key"
+          # Publish the PUBLIC half beside the cache and echo it here. Users must pin this key, so
+          # it needs to be fetchable from the cache itself rather than only from a doc that can
+          # drift — and on the first run this log line is where the value for README.md comes from.
+          # Redirect, not `| tee`: without pipefail a failing nix would be masked by tee's success
+          # and publish an EMPTY public key, which every user would then pin.
+          "$NIX" key convert-secret-to-public < "$KEYDIR/key" > nix-cache/punktfunk-cache.pub
+          cat nix-cache/punktfunk-cache.pub
+          rm -rf "$KEYDIR"
+          echo "publishing $(find nix-cache -name '*.narinfo' | wc -l) paths, $(du -sh nix-cache | cut -f1)"
+
+          # 3) Ship it. Same deploy key and retry discipline as flatpak.yml — this runner's link to
+          #    unom-1 drops TCP dials under load.
+          install -d -m700 ~/.ssh
+          printf '%s\n' "$DEPLOY_SSH_KEY" > ~/.ssh/deploy; chmod 600 ~/.ssh/deploy
+          SSH="ssh -i $HOME/.ssh/deploy -p ${DEPLOY_PORT:-22} -o StrictHostKeyChecking=accept-new"
+          DEST="${DEPLOY_USER}@${DEPLOY_HOST}"
+          bash scripts/ci/retry.sh 5 $SSH "$DEST" "mkdir -p ~/$DEPLOY_DIR/site/nar"
+          # ⚠ ORDER IS LOAD-BEARING: NARs first, narinfos second. A narinfo whose NAR has not landed
+          # yet is a HARD download failure for whoever fetches it in that window; a NAR nothing
+          # points at yet is simply invisible. rsync renames each file into place atomically, so a
+          # cancelled run (this workflow has cancel-in-progress) can only ever under-publish.
+          # No --delete: superseded paths are aged out by prune.sh below instead, so a client
+          # mid-download is never pulled out from under.
+          bash scripts/ci/retry.sh 5 rsync -az --info=stats1 -e "$SSH" nix-cache/nar/ "$DEST:$DEPLOY_DIR/site/nar/"
+          bash scripts/ci/retry.sh 5 rsync -az -e "$SSH" nix-cache/nix-cache-info nix-cache/punktfunk-cache.pub nix-cache/*.narinfo "$DEST:$DEPLOY_DIR/site/"
+          bash scripts/ci/retry.sh 5 rsync -az -e "$SSH" \
+            packaging/nix/server/compose.production.yml packaging/nix/server/Caddyfile packaging/nix/server/prune.sh \
+            "$DEST:$DEPLOY_DIR/"
+          bash scripts/ci/retry.sh 5 $SSH "$DEST" "cd ~/$DEPLOY_DIR && docker compose -f compose.production.yml up -d"
+
+          # 4) Bound it. The flatpak repo next door reached 3.84 GB publishing this same way with
+          #    no sweep, on a box that has run out of disk before; this one gets the sweep from the
+          #    first publish. Never allowed to fail the job — the cache is already live by now, and
+          #    a growing disk is a slower problem than a red main.
+          bash scripts/ci/retry.sh 3 $SSH "$DEST" "sh ~/$DEPLOY_DIR/prune.sh ~/$DEPLOY_DIR/site 180" \
+            || echo "::warning::cache prune failed — published cache may be growing unbounded"
+
+          # 5) Prove the published cache actually answers, rather than assuming the rsync landed.
+          #    A substituter that 200s on nix-cache-info but 403s on a miss is the failure mode that
+          #    breaks users' builds, so check both.
+          bash scripts/ci/retry.sh 5 curl -fsS "$CACHE_URL/nix-cache-info"
+          miss="$(curl -sS -o /dev/null -w '%{http_code}' "$CACHE_URL/0000000000000000000000000000000000.narinfo")"
+          [ "$miss" = 404 ] || { echo "::error::cache returns $miss for an absent path; nix needs 404 or every user build fails"; exit 1; }
+          echo "published → $CACHE_URL"
+
       # Opt-in only: the full Rust workspace through crane, which is the hour-long leg.
       # `github.event.inputs.*` (string) rather than `inputs.*` — the portable spelling.
       - name: Build the Rust packages (dispatch opt-in)
diff --git a/docs-site/content/docs/install.md b/docs-site/content/docs/install.md
index 8190df91..2691f8dc 100644
--- a/docs-site/content/docs/install.md
+++ b/docs-site/content/docs/install.md
@@ -32,8 +32,9 @@ On **Windows**, the host ships as a signed installer instead — see [Windows](#
 
 Each registry is public — no auth, you just trust the repo's signing key. Adding the repo is a
 one-time step covered in the linked guide; after that, normal `apt upgrade` / `dnf upgrade` /
-`pacman -Syu` (or `sudo punktfunk-sysext update` on Bazzite) tracks new builds. On **NixOS** there
-is no repo to add — you add the flake as an input and enable its module, see [NixOS](#nixos).
+`pacman -Syu` (or `sudo punktfunk-sysext update` on Bazzite) tracks new builds. On **NixOS** you add
+the flake as an input and enable its module rather than adding a package repo — but do add the
+[binary cache](#nixos), or every build compiles from source.
 
 > **Stable vs canary.** The repos in the per-distro guides are the **stable** channel — it only
 > moves when a `vX.Y.Z` release is cut. For the latest `main` build (fast, possibly broken), point
@@ -93,6 +94,22 @@ The repo's `flake.nix` is a supported install path: it builds `punktfunk-host`,
 `punktfunk-web` and `punktfunk-scripting`, and ships a NixOS module. **`x86_64-linux` only**, and
 NixOS **24.11 or newer**.
 
+**Add the binary cache first.** Without it, a build compiles the whole Rust workspace *and*
+gamescope from source — about an hour. With it you get prebuilt binaries:
+
+```nix
+nix.settings = {
+  substituters = [ "https://nix.unom.io" ];
+  trusted-public-keys = [ "punktfunk-cache-1:<key>" ];   # curl https://nix.unom.io/punktfunk-cache.pub
+};
+```
+
+Off NixOS, put the same two values in `/etc/nix/nix.conf` as `extra-substituters` /
+`extra-trusted-public-keys`. One caveat worth knowing before you copy a flake snippet from
+elsewhere: setting `inputs.punktfunk.inputs.nixpkgs.follows = "nixpkgs"` changes every store path
+and so misses the cache entirely — details in
+[packaging/nix](https://git.unom.io/unom/punktfunk/src/branch/main/packaging/nix/README.md#binary-cache-do-this-before-your-first-build).
+
 You can run it straight from the flake without NixOS (on other distros, wrap it in
 [nixGL](https://github.com/nix-community/nixGL) so the GPU drivers resolve):
 
diff --git a/packaging/nix/README.md b/packaging/nix/README.md
index 48844f00..27813187 100644
--- a/packaging/nix/README.md
+++ b/packaging/nix/README.md
@@ -45,6 +45,69 @@ GPU drivers are resolved at runtime from `/run/opengl-driver/lib`. On non-NixOS
 [nixGL](https://github.com/nix-community/nixGL) so that path is populated (`nixGL nix run …`); on
 NixOS the module (below) sets `hardware.graphics.enable = true` for you.
 
+> **Do this first, or the commands above compile the world:** without the binary cache, `nix build`
+> here means the whole Rust workspace *and* a gamescope build from source — roughly an hour on a
+> fast machine. See below.
+
+---
+
+## Binary cache (do this before your first build)
+
+CI publishes every punktfunk package to **`https://nix.unom.io`** on each push to `main` that moves
+the flake, so you get prebuilt binaries instead of an hour of `rustc`. It covers
+`punktfunk-host`, `-client`, `-tray`, `-web`, `-scripting` and `-gamescope` — everything the flake
+builds from source. Everything else in the closure is stock nixpkgs and comes from `cache.nixos.org`
+as usual, so the cache is deliberately small and adding it costs you nothing on unrelated builds.
+
+**NixOS** — in your system configuration:
+
+```nix
+nix.settings = {
+  substituters = [ "https://nix.unom.io" ];
+  trusted-public-keys = [ "punktfunk-cache-1:<PUBLIC KEY — see below>" ];
+};
+```
+
+**Anywhere else** — in `/etc/nix/nix.conf` (or `~/.config/nix/nix.conf` if you are a trusted user):
+
+```conf
+extra-substituters = https://nix.unom.io
+extra-trusted-public-keys = punktfunk-cache-1:<PUBLIC KEY — see below>
+```
+
+The current public key is served by the cache itself, so you can always check it against the source
+of truth:
+
+```sh
+curl https://nix.unom.io/punktfunk-cache.pub
+```
+
+Verify the cache is being used — this should print the store paths without compiling anything:
+
+```sh
+nix build --dry-run git+https://git.unom.io/unom/punktfunk#punktfunk-host
+```
+
+### ⚠ `nixpkgs.follows` turns the cache off
+
+Every store path is keyed by the exact inputs it was built from. Pointing punktfunk's nixpkgs at
+yours changes those inputs, so **every** path misses and you compile the workspace anyway:
+
+```nix
+# Convenient, but it costs you the entire binary cache:
+inputs.punktfunk.inputs.nixpkgs.follows = "nixpkgs";
+```
+
+That is a real trade, not a bug — `follows` buys you one shared nixpkgs in the closure instead of
+two. Take it if closure size matters more to you than build time; leave it out to get binaries.
+
+### Why not `cachix`?
+
+Nothing against it — punktfunk simply self-hosts every other channel (flatpak, deb, rpm, Arch,
+docker, winget), and a Nix cache is static files behind a web server, so it rides the same unom-1
+box and the same deploy key as the rest. Nothing about the cache is punktfunk-specific: it speaks
+plain HTTP binary-cache protocol, so any nix client works with it.
+
 ---
 
 ## NixOS module
@@ -55,6 +118,8 @@ Add the flake and enable the host and/or client:
 {
   inputs.punktfunk.url = "git+https://git.unom.io/unom/punktfunk";
   # (optional) share your nixpkgs: inputs.punktfunk.inputs.nixpkgs.follows = "nixpkgs";
+  #   ⚠ this DISABLES the binary cache — different inputs, different store paths, so every
+  #     package is rebuilt from source (~1h). See "Binary cache" above.
 
   outputs = { self, nixpkgs, punktfunk, ... }: {
     nixosConfigurations.myhost = nixpkgs.lib.nixosSystem {
@@ -369,8 +434,66 @@ RUNPATH (`/run/opengl-driver/lib`) and the GTK GApps wrapper (GSettings schemas
 are present. Fixes discovered during that bring-up: `CMAKE_POLICY_VERSION_MINIMUM=3.5` (CMake ≥ 4),
 system `libopus` (audiopus_sys), and the session Skia note above.
 
-In CI (`.gitea/workflows/nix.yml`): `nix flake check --no-build` evaluates every output *including*
-the module check above, and `punktfunk-web` + `punktfunk-scripting` are built for real. The Rust
-packages and `punktfunk-gamescope` are `workflow_dispatch` opt-ins (`build-rust`,
-`build-gamescope`) — run the latter after a `flake.lock` bump, since it patches whatever gamescope
-the pinned nixpkgs carries.
+In CI (`.gitea/workflows/nix.yml`), three tiers: `nix flake check --no-build` evaluates every output
+*including* the module check above; `punktfunk-web` + `punktfunk-scripting` are built for real on
+every PR; and on a push to `main` the Rust packages and `punktfunk-gamescope` are built and
+published to the binary cache. A `flake.lock` bump that breaks the gamescope patches therefore goes
+red on main rather than in an operator's rebuild. The `build-rust` / `build-gamescope`
+`workflow_dispatch` inputs remain, for checking those on a branch before merging.
+
+---
+
+## Cache infrastructure (maintainers)
+
+`https://nix.unom.io` is a `caddy:2-alpine` container on unom-1 serving a static directory —
+`packaging/nix/server/` — exactly like the flatpak repo (3230) and the winget source (3240). A Nix
+binary cache *is* just `nix-cache-info` + `<hash>.narinfo` + `nar/<hash>.nar.xz` behind a web
+server; there is no cache daemon to run.
+
+**Why not Gitea, and why not `storage.unom.io`:**
+
+- Gitea has 23 package registry types and none is Nix. It is not a missing label — the protocol
+  needs fixed anonymous paths at a URL *root* (`/nix-cache-info`, `/<hash>.narinfo`,
+  `/nar/…`), and `/api/packages/{owner}/generic/{name}/{version}/{file}` cannot express them.
+- The RustFS S3 at `storage.unom.io` *would* work mechanically (nix speaks `s3://…?endpoint=`, and
+  the sccache credentials already exist), but it is a local box on the home uplink with no CDN, so
+  every user download competes with CI. It also answers **403** for a missing key unless the bucket
+  policy grants anonymous `ListBucket` — and nix treats anything other than **404** as a hard error
+  rather than a cache miss, which would break users' builds for packages the cache never held.
+
+**One-time setup — in this order.** The publish step ends by fetching `nix.unom.io` to prove the
+cache really answers (and answers **404**, not 403, for a path it does not hold), so stand the
+service up *before* you set the secret that switches publishing on. The secret is the last step for
+exactly that reason: until it exists the publish no-ops with a warning and `main` stays green,
+the same way flatpak.yml's repo deploy does.
+
+1. **Edge proxy:** `nix.unom.io { reverse_proxy 192.168.50.50:3250 }` on home-reverse-proxy-1.
+2. **Port allowlist:** add `3250` to `caddy_target_ports` in `unom/infra` (proxmox/unom-1) +
+   terraform apply.
+3. **DNS:** ensure `nix.unom.io` resolves to the edge proxy.
+4. Dispatch `deploy-services.yml` (or `unom/infra`'s `deploy-all`) to bring the container up. It
+   serves an empty cache — every path 404s, which is exactly what a healthy empty cache does.
+5. Generate the signing key on a Nix box and store the secret half as the repo Actions secret
+   `NIX_CACHE_SIGNING_KEY` (the whole `name:base64` line):
+   ```sh
+   nix key generate-secret --key-name punktfunk-cache-1
+   ```
+6. Push to `main` touching the flake. The publish step prints the **public** key — paste it into the
+   "Binary cache" section above (and `docs-site/content/docs/install.md`) and commit.
+
+**Operational notes:**
+
+- Only punktfunk's own store paths are published (`nix path-info -r … | grep -- '-punktfunk'`).
+  Everything else in a closure is stock nixpkgs, already on `cache.nixos.org` behind a real CDN;
+  mirroring it would cost disk and home-to-cloud bandwidth to serve a worse copy. The publish step
+  asserts every built output is matched by that filter, so a future `pname` change fails the build
+  instead of silently dropping a package from the cache.
+- `rsync` runs **without** `--delete` (a client mid-download is never pulled out from under), and
+  NARs are uploaded *before* narinfos — a narinfo whose NAR has not landed is a hard download
+  failure for whoever fetches it in that window, while an unreferenced NAR is merely invisible.
+- Growth is bounded by `packaging/nix/server/prune.sh` (evicts narinfos untouched for 180 days,
+  then sweeps NARs nothing references). The flatpak repo next door reached 3.84 GB publishing the
+  same way with no sweep, on a box that has run out of disk before — hence the sweep from the first
+  publish. Run its self-check with `sh packaging/nix/server/prune.sh --self-test`.
+- A user on a pinned rev older than the eviction window falls back to building from source, which
+  is the pre-cache status quo.
diff --git a/packaging/nix/server/Caddyfile b/packaging/nix/server/Caddyfile
new file mode 100644
index 00000000..fc6be23f
--- /dev/null
+++ b/packaging/nix/server/Caddyfile
@@ -0,0 +1,32 @@
+# Inner Caddy (plain HTTP on :3250); the edge proxy on home-reverse-proxy-1 does TLS for
+# nix.unom.io. Serves the punktfunk Nix binary cache — which is nothing but a static tree:
+#
+#   /nix-cache-info          store dir + priority, fetched once per substituter
+#   /<32-char-hash>.narinfo  one per store path
+#   /nar/<hash>.nar.xz       the archives themselves
+#
+# ⚠ A MISSING PATH MUST 404, NOT 403. Nix reads 404 as "not in this cache, try the next
+# substituter" and treats anything else as a hard error that fails the build — so a cache
+# answering 403 for unknown hashes breaks every user who adds it, including for packages it
+# was never meant to serve. `file_server` 404s correctly; do not put an auth wrapper in front
+# of this without preserving that. This is also the concrete reason the cache is NOT a bucket
+# on storage.unom.io: S3 answers 403 for a missing key unless the bucket policy grants
+# anonymous ListBucket, and that box is on the home uplink besides.
+:3250 {
+	root * /srv
+	file_server browse
+
+	# Everything except nix-cache-info is content-addressed by the store hash and can never
+	# change meaning — a narinfo for a given hash is as immutable as the NAR it points at.
+	@immutable path /nar/* *.narinfo
+	header @immutable Cache-Control "public, max-age=31536000, immutable"
+
+	# The only mutable file, and cheap to revalidate: nix reads it once per substituter per run.
+	@info path /nix-cache-info
+	header @info Cache-Control "public, max-age=300"
+
+	# nix does not care about Content-Type, but a browser poking at the cache should not be
+	# offered a download for what is a two-line text file.
+	@text path *.narinfo /nix-cache-info
+	header @text Content-Type "text/plain; charset=utf-8"
+}
diff --git a/packaging/nix/server/compose.production.yml b/packaging/nix/server/compose.production.yml
new file mode 100644
index 00000000..770633e6
--- /dev/null
+++ b/packaging/nix/server/compose.production.yml
@@ -0,0 +1,17 @@
+# Static file server for the punktfunk Nix binary cache, on unom-1 (the DMZ services VM) —
+# the same shape as packaging/flatpak/server, because a Nix cache and an OSTree repo are both
+# just immutable files behind a web server. Caddy on home-reverse-proxy-1 terminates TLS for
+# nix.unom.io and reverse_proxies to 192.168.50.50:3250 (the port must be in unom/infra
+# caddy_target_ports). This inner Caddy serves the bind-mounted ./site tree over plain HTTP.
+# nix.yml rsyncs into ./site and runs `docker compose up -d` (idempotent).
+#
+# Port 3250: docs is 3220, flatpak 3230, winget 3240 — keep the run going.
+services:
+  nix-cache:
+    image: caddy:2-alpine
+    restart: unless-stopped
+    ports:
+      - "3250:3250"
+    volumes:
+      - ./Caddyfile:/etc/caddy/Caddyfile:ro
+      - ./site:/srv:ro
diff --git a/packaging/nix/server/prune.sh b/packaging/nix/server/prune.sh
new file mode 100755
index 00000000..f97afa92
--- /dev/null
+++ b/packaging/nix/server/prune.sh
@@ -0,0 +1,93 @@
+#!/bin/sh
+# Bound the published Nix binary cache on unom-1.
+#
+# WHY THIS EXISTS UP FRONT, rather than being added after the box fills: the flatpak repo next
+# door taught this exact lesson the expensive way. It publishes with rsync WITHOUT --delete (so
+# a client mid-download is never broken), nothing ever removed the superseded objects, and it
+# reached 3.84 GB on a box that had already run out of disk once. This cache has the same
+# publish model and the same growth shape — every build whose inputs moved adds a fresh set of
+# store paths and keeps the old ones — so it gets the sweep from day one.
+#
+# The order below is the whole correctness argument:
+#
+#   1. Delete narinfos older than $DAYS. rsync -a carries the CI-side mtime over, and `nix copy`
+#      rewrites every narinfo it publishes on every run, so a path that is still being published
+#      keeps getting a fresh mtime. Age therefore means "no publish has referenced this in
+#      $DAYS", which is exactly the eviction signal wanted.
+#   2. THEN delete NARs no surviving narinfo points at.
+#
+# Doing it the other way round — or aging the NARs independently — can strand a live narinfo
+# pointing at a deleted NAR, and that is strictly worse than a cache miss: nix reports a missing
+# NAR as a hard download failure, not as "not cached, build it yourself".
+#
+# POSIX sh: this runs over ssh on unom-1 (Debian), invoked by .gitea/workflows/nix.yml.
+#
+# Usage:  sh prune.sh <cache-dir> [max-age-days]     (default 180)
+#         sh prune.sh --self-test
+set -eu
+
+self_test() {
+	# Smallest thing that fails if the ordering or the reference sweep breaks.
+	t="$(mktemp -d)"
+	trap 'rm -rf "$t"' EXIT
+	mkdir -p "$t/nar"
+	printf 'StoreDir: /nix/store\nWantMassQuery: 1\nPriority: 41\n' >"$t/nix-cache-info"
+
+	# A live path, a stale one, and a NAR nothing ever pointed at.
+	printf 'StorePath: /nix/store/aaa-live\nURL: nar/live.nar.xz\n' >"$t/aaa.narinfo"
+	printf 'StorePath: /nix/store/bbb-stale\nURL: nar/stale.nar.xz\n' >"$t/bbb.narinfo"
+	: >"$t/nar/live.nar.xz"
+	: >"$t/nar/stale.nar.xz"
+	: >"$t/nar/orphan.nar.xz"
+	# Two paths that SHARE a NAR, one stale and one live: the shared NAR must survive. A sweep
+	# that deleted NARs per-evicted-narinfo instead of by surviving references would drop it.
+	printf 'StorePath: /nix/store/ccc-live\nURL: nar/shared.nar.xz\n' >"$t/ccc.narinfo"
+	printf 'StorePath: /nix/store/ddd-stale\nURL: nar/shared.nar.xz\n' >"$t/ddd.narinfo"
+	: >"$t/nar/shared.nar.xz"
+
+	# Age the stale ones well past any plausible threshold (portable -t form: YYYYMMDDhhmm).
+	touch -t 200001010000 "$t/bbb.narinfo" "$t/ddd.narinfo" "$t/nar/stale.nar.xz"
+
+	prune "$t" 180
+
+	fail() { echo "SELF-TEST FAILED: $1" >&2; exit 1; }
+	[ -f "$t/aaa.narinfo" ] || fail "evicted a fresh narinfo"
+	[ -f "$t/nar/live.nar.xz" ] || fail "evicted a referenced NAR"
+	[ -f "$t/nix-cache-info" ] || fail "deleted nix-cache-info"
+	[ ! -f "$t/bbb.narinfo" ] || fail "kept a stale narinfo"
+	[ ! -f "$t/nar/stale.nar.xz" ] || fail "kept a NAR nothing references any more"
+	[ ! -f "$t/nar/orphan.nar.xz" ] || fail "kept an orphan NAR"
+	[ -f "$t/nar/shared.nar.xz" ] || fail "deleted a NAR a surviving narinfo still references"
+	echo "prune.sh self-test OK"
+}
+
+prune() {
+	root="$1"
+	days="$2"
+	cd "$root"
+
+	before="$(du -sh . 2>/dev/null | cut -f1)"
+
+	# 1. Age out the narinfos.
+	find . -maxdepth 1 -name '*.narinfo' -mtime "+$days" -delete
+
+	# 2. Sweep NARs nothing points at any more. Both lists are relative to $root and spelled the
+	#    same way ("nar/<file>") so `comm` can diff them.
+	keep="$(mktemp)"
+	have="$(mktemp)"
+	# `|| true`: an empty cache (or one whose narinfos were all just evicted) makes the glob
+	# match nothing, and an empty keep-list is the correct answer there, not an error.
+	cat ./*.narinfo 2>/dev/null | sed -n 's|^URL: ||p' | sort -u >"$keep" || true
+	find nar -type f 2>/dev/null | sed 's|^\./||' | sort >"$have" || true
+	comm -13 "$keep" "$have" | tr '\n' '\0' | xargs -0 -r rm -f
+	rm -f "$keep" "$have"
+
+	# Leave the numbers in the deploy log — this is the only place the published size is visible.
+	echo "cache pruned (narinfos older than ${days}d): ${before:-?} -> $(du -sh . 2>/dev/null | cut -f1) in $(pwd)"
+}
+
+case "${1:-}" in
+--self-test) self_test ;;
+"") echo "usage: prune.sh <cache-dir> [max-age-days] | --self-test" >&2; exit 2 ;;
+*) prune "$1" "${2:-180}" ;;
+esac