diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index ca89813..537076f 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -40,6 +40,9 @@ jobs:
       - name: Test (unit + property + worker + type-level, coverage-gated)
         run: pnpm test -- --coverage
 
+      - name: Bundle-size budget (tree-shaken min+gzip)
+        run: pnpm size
+
   # ---------------------------------------------------------------------------
   # Runtime lanes (P3): prove the SHIPPED dist actually runs on each claimed runtime.
   # Each lane builds first (the smoke imports packages/ecsia/dist), then runs the SAME
diff --git a/bundle-budget.json b/bundle-budget.json
new file mode 100644
index 0000000..4bdc86d
--- /dev/null
+++ b/bundle-budget.json
@@ -0,0 +1,17 @@
+{
+  "_comment": "Bundle-size budgets: max min+gzip BYTES per tree-shaken entry (scripts/size-check.mjs). The honest \"lean install\" number. Ratchet DOWN with `node scripts/size-check.mjs --update` when a build shrinks; CI fails if a build grows past budget*1.03.",
+  "budgets": {
+    "kernel": {
+      "gzip": 39987,
+      "min": 126142
+    },
+    "core-min": {
+      "gzip": 30154,
+      "min": 97265
+    },
+    "full-umbrella": {
+      "gzip": 54587,
+      "min": 171999
+    }
+  }
+}
diff --git a/package.json b/package.json
index bbd28ad..1e2ebc6 100644
--- a/package.json
+++ b/package.json
@@ -21,6 +21,8 @@
     "smoke:runtime": "node scripts/runtime-smoke.mjs",
     "smoke:browser:bundle": "node scripts/browser-smoke/build.mjs",
     "smoke:browser": "playwright test --config scripts/browser-smoke/playwright.config.ts",
+    "size": "node scripts/size-check.mjs",
+    "size:update": "node scripts/size-check.mjs --update",
     "docs:api": "typedoc",
     "docs:check": "node scripts/check-doc-snippets.mjs",
     "docs:dev": "vitepress dev website",
diff --git a/scripts/size-check.mjs b/scripts/size-check.mjs
new file mode 100644
index 0000000..2dc4bc4
--- /dev/null
+++ b/scripts/size-check.mjs
@@ -0,0 +1,119 @@
+// Bundle-size budget. Bundles representative tree-shaken entry points against the BUILT dist,
+// minifies (esbuild) + gzips (zlib), and asserts each stays under a committed budget. This is the
+// honest "lean install" number — what a bundler ships for a real app, NOT the unminified source.
+// Run after `pnpm build`. `--update` rewrites the budget to the measured sizes (the ratchet).
+//
+// Why per-import, not one number: ecsia is batteries-included but `sideEffects: false`, so a typed
+// data + scheduler app pulls only the kernel; relations / serialization / topics drop unless used.
+// The budgets pin that tree-shaking stays honest.
+
+import { build } from 'esbuild'
+import { gzipSync } from 'node:zlib'
+import { readFileSync, writeFileSync, existsSync } from 'node:fs'
+import { fileURLToPath } from 'node:url'
+import { dirname, resolve } from 'node:path'
+
+const HERE = dirname(fileURLToPath(import.meta.url))
+const ROOT = resolve(HERE, '..')
+const BUDGET_PATH = resolve(ROOT, 'bundle-budget.json')
+
+// Each entry is the smallest realistic import for a use-case. Measured tree-shaken: the umbrella is
+// pure static re-exports, so unused subsystems drop.
+const ENTRIES = [
+  {
+    name: 'kernel',
+    note: 'typed data + systems + scheduler (the typical app)',
+    code: `import { createWorld, defineComponent, defineSystem, createScheduler, read, write } from 'ecsia'
+           console.log(createWorld, defineComponent, defineSystem, createScheduler, read, write)`,
+  },
+  {
+    name: 'core-min',
+    note: 'just a world + a component (@ecsia/core alone)',
+    code: `import { createWorld, defineComponent } from '@ecsia/core'
+           console.log(createWorld, defineComponent)`,
+  },
+  {
+    name: 'full-umbrella',
+    note: 'everything re-exported from the umbrella (the upper bound)',
+    code: `import * as ecsia from 'ecsia'
+           console.log(ecsia)`,
+  },
+]
+
+const alias = {
+  ecsia: resolve(ROOT, 'packages/ecsia/dist/index.js'),
+  '@ecsia/core': resolve(ROOT, 'packages/core/dist/index.js'),
+  '@ecsia/schema': resolve(ROOT, 'packages/schema/dist/index.js'),
+  '@ecsia/scheduler': resolve(ROOT, 'packages/scheduler/dist/index.js'),
+  '@ecsia/relations': resolve(ROOT, 'packages/relations/dist/index.js'),
+  '@ecsia/serialization': resolve(ROOT, 'packages/serialization/dist/index.js'),
+}
+
+async function measure(entry) {
+  const out = await build({
+    stdin: { contents: entry.code, resolveDir: ROOT, loader: 'js' },
+    bundle: true,
+    minify: true,
+    format: 'esm',
+    platform: 'neutral',
+    write: false,
+    legalComments: 'none',
+    alias,
+    // node:worker_threads is only reached on the threaded path via dynamic import — never in a
+    // browser bundle; mark it external so it doesn't inflate the measured size.
+    external: ['node:worker_threads', 'node:url'],
+  })
+  const code = out.outputFiles[0].contents
+  return { min: code.length, gzip: gzipSync(code).length }
+}
+
+async function main() {
+  for (const a of Object.values(alias)) {
+    if (!existsSync(a)) {
+      console.error(`size-check: missing build artifact ${a} — run \`pnpm build\` first.`)
+      process.exit(1)
+    }
+  }
+  const update = process.argv.includes('--update')
+  const budget = existsSync(BUDGET_PATH)
+    ? JSON.parse(readFileSync(BUDGET_PATH, 'utf8'))
+    : { _comment: '', budgets: {} }
+
+  const results = {}
+  let failed = false
+  const TOLERANCE = 1.03 // 3% headroom so a trivial change doesn't trip the gate; ratchet with --update
+  for (const entry of ENTRIES) {
+    const { min, gzip } = await measure(entry)
+    results[entry.name] = { gzip, min }
+    const limit = budget.budgets?.[entry.name]?.gzip
+    const status = limit === undefined ? 'NEW' : gzip <= Math.ceil(limit * TOLERANCE) ? 'ok' : 'OVER'
+    if (status === 'OVER') failed = true
+    console.log(
+      `  ${entry.name.padEnd(16)} ${String(gzip).padStart(6)} B gz  (${String(min).padStart(6)} B min)  ` +
+        `${limit === undefined ? '(no budget)' : `budget ${limit} B`}  ${status}` +
+        `   — ${entry.note}`,
+    )
+  }
+
+  if (update) {
+    const next = {
+      _comment:
+        'Bundle-size budgets: max min+gzip BYTES per tree-shaken entry (scripts/size-check.mjs). The honest "lean install" number. Ratchet DOWN with `node scripts/size-check.mjs --update` when a build shrinks; CI fails if a build grows past budget*1.03.',
+      budgets: Object.fromEntries(Object.entries(results).map(([k, v]) => [k, { gzip: v.gzip, min: v.min }])),
+    }
+    writeFileSync(BUDGET_PATH, JSON.stringify(next, null, 2) + '\n')
+    console.log(`\nsize-check: budget written to ${BUDGET_PATH}`)
+    return
+  }
+
+  if (failed) {
+    console.error('\nsize-check: FAILED — a bundle grew past its budget. Investigate, or `--update` to ratchet if intended.')
+    process.exit(1)
+  }
+  console.log('\nsize-check: OK — all bundles within budget.')
+}
+
+main().catch((err) => {
+  console.error(err)
+  process.exit(1)
+})
diff --git a/website/guide/performance.md b/website/guide/performance.md
index e458526..3fef9d8 100644
--- a/website/guide/performance.md
+++ b/website/guide/performance.md
@@ -179,14 +179,23 @@ pnpm bench:macro:pool   # the worker-pool speedup sweep on its own, with the pri
 
 ## Regression guard
 
-CI does **not** run `bench:report`. Wall-clock time on shared CI runners is noise — a slow neighbor on
-the runner would flap the suite — so we deliberately **do not assert milliseconds in CI**. What CI
-*does* guard is correctness and behavior, with counter-based assertions: the worker-pool smoke test (a
-quick check that it compiles and runs, not a measurement) asserts every threaded configuration is
-byte-identical to single-thread (the `byte-identical` column above), and the bench builders are
-cross-checked so neither `eachChunk` nor `bindColumns` can silently diverge from `.each`. Performance
-regressions are caught by re-running `bench:report` on a fixed machine and comparing `RESULTS.json`,
-not by a CI timer.
+CI does **not** assert milliseconds — wall-clock time on a shared runner is noise. It guards
+performance two ways instead. First, correctness: the worker-pool smoke test asserts every threaded
+configuration is byte-identical to single-thread (the `byte-identical` column above), and the bench
+builders are cross-checked so neither `eachChunk` nor `bindColumns` can silently diverge from `.each`.
+Second, a dedicated bench job asserts each iteration path's ns/entity **ratio against a same-run bitECS
+control** stays under a committed ceiling — the ratio cancels runner drift, so it catches a real
+regression (say, `bindColumns` deopting) without flapping on a slow neighbor. The absolute published
+numbers still come from `bench:report` on a fixed machine.
+
+## Bundle size
+
+The kernel — typed data, systems, queries, and the scheduler — is about **40 KB min+gzip**; just a
+world and components (`@ecsia/core` alone) is ~30 KB, and importing *everything* from the umbrella is
+~55 KB. ecsia is batteries-included, so it is larger than a minimal core like bitECS (~5 KB); the
+packages are `sideEffects: false`, so a bundler ships only the subsystems you import — relations,
+serialization, and topics drop unless used. A CI budget (`pnpm size`) holds these numbers in place so
+a change can't quietly inflate the install.
 
 ## See also