diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index a73c78e8..c371679b 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -177,23 +177,23 @@ jobs:
         with:
           targets: ${{ matrix.rust_target }}
 
-      # Release builds compile the same dependency graph the test workflows
-      # do, from scratch, on four runners. The measurement that motivated
-      # caching is in test.yml: 7m57s of a 12m14s job was the release compile
-      # alone.
+      # No cargo cache here, deliberately, and it is a budget decision rather
+      # than a judgement that caching does not work.
       #
-      # `workspaces` names the cargo workspace root; without it the action
-      # looks for Cargo.lock at the repository root and caches nothing.
+      # A cargo cache is 1.0-1.2 GB per entry. This workflow has four matrix
+      # entries, so caching it would claim another ~4.5 GB of a 10 GB
+      # per-repository limit that already holds ~110 npm caches and the three
+      # entries test_build.yml writes on master. Over that limit GitHub evicts
+      # least-recently-used, and these would be the least recently used of all:
+      # this workflow is `workflow_dispatch` only and runs about once a week.
       #
-      # `key` is not optional here. The automatic key covers the job id, the
-      # rustc release/host and a hash of the Cargo files -- and the two
-      # windows-latest entries share all three, differing only in `arch`. The
-      # ARM64 job would otherwise restore the x64 job's target directory.
-      - name: Cache Rust dependencies
-        uses: Swatinem/rust-cache@v2
-        with:
-          workspaces: src-tauri
-          key: ${{ matrix.os }}-${{ matrix.arch }}
+      # What it would buy is also the least valuable minutes in the repository.
+      # A release build is not on anyone's critical path -- nobody waits on it
+      # the way they wait on a pull request check -- and its four jobs already
+      # finish in 12-16 minutes. Spending a third of the cache budget to make a
+      # weekly job faster, at the cost of the caches every pull request reads,
+      # is the wrong trade. If the limit ever stops binding, this is the first
+      # place to add one back.
 
       - name: Install Linux dependencies
         if: matrix.platform == 'ubuntu-24.04'
diff --git a/.github/workflows/test.yml b/.github/workflows/test.yml
index 62791051..75ae4a86 100644
--- a/.github/workflows/test.yml
+++ b/.github/workflows/test.yml
@@ -36,11 +36,27 @@ jobs:
     steps:
       - uses: actions/checkout@v7
 
+      # No `cache: npm`, deliberately — it was the single largest consumer of a
+      # cache budget the cargo caches above have to fit inside.
+      #
+      # Measured: 113 npm cache entries, 8.76 GB, spread over 50 refs, against
+      # a 10 GB per-repository limit. GitHub scopes caches by ref, so one is
+      # written per branch per platform and nothing can bound it the way
+      # `save-if` bounds the cargo caches. Most of those 113 belonged to pull
+      # requests closed weeks earlier; caches are only evicted after seven days
+      # unused, or by least-recently-used pressure that takes the biggest
+      # entries — the cargo caches — first.
+      #
+      # What it bought, measured on runs that hit it: `npm ci` in 6s on Linux,
+      # 9s on macOS, 15-17s on Windows. It caches `~/.npm`, the download
+      # directory, not `node_modules` — `npm ci` still runs and still links all
+      # 190 packages, so the saving is only the registry fetch, and the runners
+      # are close to the registry. Seconds per job, for the budget that makes
+      # an eight-minute saving possible.
       - name: setup node
         uses: actions/setup-node@v7
         with:
           node-version: lts/*
-          cache: 'npm'
 
       - name: install rust stable
         uses: dtolnay/rust-toolchain@stable
@@ -55,14 +71,21 @@ jobs:
       # looks for Cargo.lock at the repository root, finds none and caches
       # nothing.
       #
-      # This job is also what makes the cache reach anything: a pull request
-      # can only restore caches written on its own branch or on the base
-      # branch, and of the three workflows only this one runs on push to
-      # master. Its master runs are what seed a freshly opened branch.
+      # A pull request can only restore caches written on its own branch or on
+      # the base branch, so master runs are what seed a freshly opened branch.
+      # This job and test_build.yml's matrix both run on push to master for
+      # that reason, and `save-if` keeps them the only writers: a pull request
+      # restores and saves nothing.
+      #
+      # That bound is not cosmetic. One run's cargo caches measured 4.45 GB
+      # against a 10 GB per-repository limit shared with ~110 npm caches, and
+      # GitHub evicts least-recently-used -- so unbounded per-branch copies
+      # would push out the very entries they are meant to reuse.
       - name: cache rust dependencies
         uses: Swatinem/rust-cache@v2
         with:
           workspaces: src-tauri
+          save-if: ${{ github.ref == 'refs/heads/master' }}
 
       - name: install dependencies (ubuntu only)
         run: |
diff --git a/.github/workflows/test_build.yml b/.github/workflows/test_build.yml
index 04a37669..553cd503 100644
--- a/.github/workflows/test_build.yml
+++ b/.github/workflows/test_build.yml
@@ -27,15 +27,39 @@ on:
       - '**.md'
       - 'pics/**'
       - '.github/ISSUE_TEMPLATE/**'
+  # This trigger exists to fill the cargo cache, not to test master again.
+  #
+  # A GitHub Actions cache can only be restored by the branch that wrote it or
+  # by a branch descended from one that did, and until now nothing wrote a
+  # cargo cache on master: this workflow only ran on `pull_request`, so every
+  # cache was scoped to a pull request branch and no other branch could ever
+  # read it. Every new pull request paid a full cold compile no matter how many
+  # had been cached before it -- which is most of the benefit of caching, lost
+  # to where the cache was written rather than to anything about caching.
+  #
+  # One run per merge seeds all three platforms for every pull request opened
+  # afterwards. Only the first is a cold compile; later master runs restore
+  # master's own cache and recompile little.
+  push:
+    branches:
+      - master
+    paths-ignore:
+      - '**.md'
+      - 'pics/**'
+      - '.github/ISSUE_TEMPLATE/**'
   workflow_dispatch:
 
 # Every push to a pull request queued another full matrix while the previous
 # one was still waiting for a runner, and both then competed for the same
 # scarce macOS capacity. Only the newest revision of a branch is worth
 # building.
+#
+# Not on master, for the same reason test.yml does not: the push trigger above
+# exists to write the cache, and cancelling one master run because a second
+# merge landed would throw away the compile that was going to fill it.
 concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}
-  cancel-in-progress: true
+  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
 
 jobs:
   build-test:
@@ -64,11 +88,12 @@ jobs:
     steps:
       - uses: actions/checkout@v7
 
+      # No `cache: npm` — see the note in test.yml. It cost 8.76 GB of a 10 GB
+      # budget across 113 unbounded per-branch entries and saved seconds.
       - name: setup node
         uses: actions/setup-node@v7
         with:
           node-version: lts/*
-          cache: npm
 
       - name: install rust stable
         uses: dtolnay/rust-toolchain@stable
@@ -89,16 +114,23 @@ jobs:
       # `key` makes the cache say which matrix entry wrote it rather than
       # leaning on that coincidence.
       #
-      # Note what this cannot do on its own: this workflow has no push
-      # trigger, so its caches are only ever written on pull request branches
-      # and a branch can only restore its own or the base branch's. The first
-      # run of a new pull request will still be a cold compile; later pushes
-      # to the same branch are the ones that hit.
+      # `save-if` is what keeps this affordable. Without it every pull request
+      # writes its own copy, and a copy here is 1.0-1.2 GB per platform: four
+      # cargo caches from a single run measured 4.45 GB, against a 10 GB
+      # per-repository limit that also holds ~110 npm caches. Two pull requests
+      # in flight would overflow it, and GitHub evicts least-recently-used --
+      # so the largest entries, these, are the ones that vanish, and the cache
+      # would be a slower way of not caching.
+      #
+      # Writing only on master pins the total at one set. Pull requests restore
+      # from master and save nothing, which is also what makes the push trigger
+      # above worth its runner time: it is the only writer.
       - name: cache rust dependencies
         uses: Swatinem/rust-cache@v2
         with:
           workspaces: src-tauri
           key: ${{ matrix.os-name }}
+          save-if: ${{ github.ref == 'refs/heads/master' }}
 
       - name: install dependencies (ubuntu only)
         if: matrix.platform == 'ubuntu-22.04'
diff --git a/scripts/ciCacheBudget.test.ts b/scripts/ciCacheBudget.test.ts
new file mode 100644
index 00000000..49615255
--- /dev/null
+++ b/scripts/ciCacheBudget.test.ts
@@ -0,0 +1,56 @@
+import assert from 'node:assert/strict';
+import test from 'node:test';
+
+import { readSource } from './sourceTree.js';
+
+const workflows = ['build.yml', 'test.yml', 'test_build.yml', 'publish-packages.yml'].map(
+	(name) => ({ name, text: readSource(`.github/workflows/${name}`) }),
+);
+
+// GitHub gives a repository 10 GB of Actions cache and evicts
+// least-recently-used over it. Both assertions below exist because that budget
+// was measured being exceeded — 13.22 GB across 117 entries — and because the
+// entries that overflow it are evicted in the order that hurts most: the
+// largest first, which is exactly the cargo caches the budget is being spent
+// on. Neither of these is a law of nature; both are trades with numbers behind
+// them, recorded in the workflow comments. Change them by re-doing the
+// measurement, not by deleting the assertion.
+
+test('a cargo cache is written on master and nowhere else', () => {
+	// Without `save-if` every pull request writes its own copy, and a copy is
+	// 1.0-1.2 GB per platform: one run of #572 wrote 4.45 GB. Two pull requests
+	// in flight overflow the budget on their own, and what gets evicted is the
+	// cargo caches — so the cache becomes a slower way of not caching.
+	//
+	// This is also why test_build.yml runs on push to master. The two are one
+	// decision: `save-if` makes master the only writer, so master has to run.
+	for (const { name, text } of workflows) {
+		const uses = [...text.matchAll(/uses: Swatinem\/rust-cache@v2\n([\s\S]{0,400}?)(?=\n\s*- name:|\n\n)/g)];
+		for (const [, block] of uses) {
+			assert.match(
+				block,
+				/save-if: \$\{\{ github\.ref == 'refs\/heads\/master' \}\}/,
+				`${name} caches cargo output without restricting the save to master`,
+			);
+		}
+	}
+});
+
+test('setup-node does not cache npm', () => {
+	// It was the largest consumer of the same budget: 113 entries, 8.76 GB,
+	// across 50 refs, most of them pull requests closed weeks earlier. Caches
+	// are scoped by ref, so one is written per branch per platform and nothing
+	// can bound it the way `save-if` bounds the cargo caches.
+	//
+	// What it bought, on runs that hit it: `npm ci` in 6s on Linux, 9s on
+	// macOS, 15-17s on Windows. It caches `~/.npm`, not `node_modules`, so
+	// `npm ci` still runs and still links every package — the saving is the
+	// registry fetch alone.
+	for (const { name, text } of workflows) {
+		assert.doesNotMatch(
+			text,
+			/^\s*cache: '?npm'?\s*$/m,
+			`${name} re-enables the npm cache; re-read the budget note in test.yml first`,
+		);
+	}
+});