Native Node ignores ts-node's resolver: it reads package.json exports
and will neither fall back to index.js nor extension-resolve bare
specifiers. Add an exports map pointing at the .ts sources so both
native Node and the host's Vite/Embroider build resolve the package
and its subpaths (bare entry, explicit directory-index subpaths, and a
"./*" -> "./*.ts" wildcard); keep the existing #fetch/#lint-task
imports. Do not set "type":"module" — the package ships CJS .js eslint
configs that would break, so Node's per-file syntax detection handles
the mix.

Replace the runtime constructs native Node rejects under ESM: the lazy
require() in fetch-node.ts (now createRequire) and url-signature.ts
(now a bare crypto import, which the host already aliases to
crypto-browserify), and __dirname -> import.meta.dirname. Flip the
bench:amd scripts and their dynamic requires to native node, drop the
ts-node devDependency, and drop the now-redundant .ts extension on the
marked-sync shim import in the host so it resolves through the exports
map.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…e ESM

The host's Vite/Embroider build adds extensions and index-resolves, so it
hid import forms that native Node's ESM resolver rejects. Make every
import in runtime-common resolvable by native Node (verified by importing
the package and all subpaths under `node`, no bundler):

- Add `.js` to extensionless CJS-package subpath imports that have no
  exports map (lodash/<method>, matrix-js-sdk/lib/utils — the latter has
  both a utils.js file and a utils/ directory, and ESM picks the dir).
- Replace the named import from lodash's CJS root in catalog.ts with
  per-method default imports (Node can't statically detect lodash's CJS
  named exports).
- Point the three bare-directory barrel imports (`from '.'`) at
  `./index.ts` (native ESM does not index-resolve directories).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

postgres is the first leaf consumer of runtime-common to run on native
Node. Add an exports map pointing at its .ts sources (".", "./*", and
"./package.json") so consumers resolve the package and its subpaths
without ts-node; replace __dirname with import.meta.dirname in
pg-adapter.ts and convert-to-sqlite.ts; flip the migrate and make-schema
scripts from ts-node to native node; and drop the ts-node dependency.
Leave "type" unset so Node's syntax detection keeps the package's CJS
.js migration helpers working alongside the ESM .ts sources.

Verified by importing @cardstack/postgres (and its runtime-common
dependency chain) under native node.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Foundation toward the full node-run ESM migration: type:module on
runtime-common + postgres, nested package.json {type:commonjs} for their
CJS dirs (etc/eslint, migrations), .cjs renames for stray CJS files, and
revert of the collateral host/eslint-plugin engines bumps. Resolves
TS1470 but TS2307 (base path aliases) and TS2349 (CJS interop) remain,
plus all consumers still need migrating. Parked — not green.

Swap ts-node for native Node (>=24) TypeScript execution across the
node-run package cluster: runtime-common, postgres, billing,
realm-server, realm-test-harness, ai-bot, bot-runner, matrix, and
software-factory. All five service entries plus ai-bot and bot-runner
now link their full module graph under native node.

Native Node ESM is stricter than ts-node/Vite about resolution; the
recurring breakages and their fixes are captured as a reusable,
idempotent codemod under scripts/esm-codemod/ (run.mjs orchestrator,
per-rule modules, and README.md documenting the full error taxonomy):

- add explicit extensions to relative imports (no extension search)
- switch lodash -> lodash-es named imports repo-wide
- CJS named imports (fs-extra, debug) -> default + destructure
- __dirname/__filename -> import.meta.dirname/filename
- ts-node --transpileOnly <entry> -> node <entry>.ts in
  package.json / *.sh / mise-tasks

Manual fixes not suited to a blind codemod: exports maps + type:module
on billing and realm-test-harness; in-source spawn('ts-node', ...) ->
spawn('node', ['x.ts', ...]); postgres .js->.cjs import; matrix-js-sdk
deep-import .js suffixes; and an import.meta.url-as-path bug in lint.ts.

Still remaining: the qunit test-runner bootstrap (qunit --require
ts-node/register -> node tests/index.ts) and removing the ts-node
devDependency afterward.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Replace `qunit --require ts-node/register/transpile-only tests/index.ts`
with `node tests/index.ts` for ai-bot, bot-runner, software-factory, and
realm-server. The qunit CLI used to supply the QUnit global, TAP reporter,
autostart, and a failure-based exit code; each package now has a
tests/qunit-bootstrap.ts that wires those up (autostart off, TAP reporter,
runEnd exit code) and its tests/index.ts ends with QUnit.start().

Test files imported `{ module, test }` from qunit, which Node's ESM loader
can't read as named exports from the CJS package; the cjs-named-to-default
codemod (now line-anchored so it skips qunit imports embedded in
test-fixture strings) rewrites them to a default import + destructure.
Also adds jsonwebtoken to that codemod's CJS list.

realm-server's tests/index.ts additionally needed a createRequire shim so
its synchronous, order-preserving test-file loader keeps working, with an
explicit `.ts` on each require (native require does no extension search);
its CI JUnit reporter is renamed .cjs (realm-server is now type:module).

Removes the now-unused ts-node devDependency from every package and fixes
two `#!/usr/bin/env ts-node` shebangs and a `ps | grep ts-node`
process-matcher that no longer matches the node-run children.

Known remaining: ai-bot has 5 tests that reassign sealed ES-module
namespace bindings for mocking (needs DI); software-factory's full node
suite needs boxel-cli and its own src migrated off bare require().

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Two ESM-semantic issues surfaced once ai-bot ran under `node tests/index.ts`:

1. node-pg-migrate default-export interop (postgres). `import migrate from
   'node-pg-migrate'` binds the whole CJS namespace under native ESM, so the
   runner (on `.default`) wasn't callable — "migrate is not a function". This
   broke DB migration everywhere, not just tests. Resolve `.default ?? ns`. Also
   add `package.json` to the migrate `ignorePattern` so the `type:commonjs`
   marker file in the migrations dir isn't loaded as a migration.

2. Sentry mocking against a sealed ES-module namespace. Tests reassigned
   `(Sentry as any).captureException`, a no-op because `import * as Sentry`
   bindings are read-only. Route the source through a mutable `lib/sentry.ts`
   `errorReporter` indirection that responder.ts/debug.ts call and the tests
   stub.

ai-bot now passes 170/170 (locking tests need the usual PG env vars). Documents
both as error classes #12 and #13 in the codemod README.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

software-factory's node test suite crashed at boot on `require('@cardstack/logger')`
in src/logger.ts — `require` doesn't exist in ESM scope. @cardstack/logger is a
CJS package that ships no type declarations, so the existing code cast the
`require()` result; keep that by defining `require` via createRequire rather than
switching to a static import (which would trip TS7016). Same fix for the
identical realm-test-harness/src/logger.ts.

With this and the previously-added @cardstack/boxel-cli `exports` map, the
software-factory node suite runs: 452/453 pass. The one failure
(port-allocator dual-stack `::` bind) is a macOS-vs-Linux socket-semantics
difference, not a migration regression — it asserts Linux behavior and fails the
same way under ts-node on macOS.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Packages flipped to "type":"module" (billing, realm-server, realm-test-harness)
treat their `.eslintrc.js` as ESM, so ESLint fails to load them
("module is not defined in ES module scope"). Rename those configs to
`.eslintrc.cjs`, which ESLint still auto-discovers.

Same problem for realm-server's CJS helper scripts run via `node`:
`shard-test-modules.js` (invoked in CI to compute test shards) and
`run-test-modules.js` both use `require`. Rename to `.cjs` and update the CI
workflow reference.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…odemod README

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

setup-localhost-resolver.ts (undici/dns, gated on BOXEL_ENVIRONMENT) and
lib/wtfnode-on-signal.ts (wtfnode, gated on BOXEL_WTFNODE) call `require`,
which doesn't exist in ESM scope — they'd throw the moment their gate is set.
Recreate `require` via createRequire inside each guard, preserving the lazy,
optional-dependency semantics (matching the existing fetch-node.ts pattern).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Flipping the node cluster to "type":"module" switches ember-tsc's nodenext
resolution from CJS-mode to ESM-mode, which surfaced the parked-WIP type errors.
Root causes and fixes:

- Base-realm path alias (the bulk): under ESM-mode nodenext the extensionless
  `"https://cardstack.com/base/*": ["../base/*"]` mapping no longer resolves
  (.gts/.ts need explicit extensions). Expand every tsconfig's mapping to try
  `*.gts`/`*.ts`/`*.d.ts`/`*`. This also clears the transitive TS2307 cascade in
  consumers (catalog, host, …) that type-check runtime-common's now-ESM sources.
- CJS default-interop the nodenext type layer can't model even though it works at
  runtime: bump magic-string 0.25.9 -> ^0.30.21 (properly packaged exports+types;
  AMD transpiler output verified unchanged); re-export `ignore` once with its call
  signature (runtime-common/ignore.ts); type node-pg-migrate's runner via
  RunnerOption.
- Make ai-bot/bot-runner/software-factory/matrix "type":"module" (they already run
  as ESM and use import.meta, which CJS-mode rejected with TS1470); switch matrix's
  tsconfig to nodenext + skipLibCheck; rename their CJS .eslintrc.js -> .cjs.
- Misc: `import type` from a bare `utils/jwt` -> relative; matrix-js-sdk type
  deep-import gains .js; QUnit.reporters/on cast (missing from @types/qunit);
  drop a now-conflicting `declare const QUnit`; sweep 3 catalog cards to lodash-es.

All nine node-cluster packages now type-check clean (0 errors): runtime-common,
postgres, billing, realm-server, realm-test-harness, ai-bot, bot-runner,
software-factory, matrix. Suites still pass (ai-bot 170, bot-runner 38) and the
service entries still load.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…od README

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

CI surfaced two gaps from the lodash-es switch and the require shims:

- base, boxel-ui, and experiments-realm import lodash-es but never declared it
  (the dep swap only covered packages that previously declared `lodash`), so
  their builds/type-checks failed to resolve the module. Add `lodash-es`
  (boxel-ui also swaps `@types/lodash` -> `@types/lodash-es`).
- realm-server lint:js: prettier-wrap the multi-line CJS destructures the
  interop codemod emitted, and drop the now-unused
  `eslint-disable @typescript-eslint/no-var-requires` directives left above the
  createRequire-shimmed `require()` calls.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Two more CI failures from the native-ESM move:

- Build / Realm Performance Benchmark: the host build's getLatestSchemaFile()
  (config/environment.js) picked the lexically-last entry in the migrations and
  schema dirs. The migrations dir now contains `package.json` (pins it to
  type:commonjs), which sorts after the timestamped migrations and made the
  freshness check compare `package.json` against the schema timestamp → false
  "schema out of date". Filter both lists to timestamped files only.
- AMD Transpile Bench: `if (require.main === module)` entry-point guards throw
  `require is not defined` under ESM. Replace with `import.meta.main` (Node 24)
  across the affected scripts in runtime-common, realm-server, matrix, and
  boxel-cli.

AMD bench gate passes locally (also confirms magic-string 0.30 transpile output
is within the perf tolerance).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…uild

boxel-cli is a published CJS package (esbuild emits format:cjs), but the
ts-node->node conversion of its dev/build scripts made node run them as ESM
(import syntax + type:unset), so their `__dirname`/`require.main` broke:

- Convert the build scripts (build.ts, build-types.ts, build-test-harness.ts,
  build-realms.ts, build-skills.ts, build-plugin.ts) to import.meta.dirname /
  import.meta.main. Leave src/ on __dirname — esbuild bundles it to CJS, where
  __dirname is correct and import.meta would be empty.
- tsconfig -> module:esnext + moduleResolution:bundler so the now-ESM scripts'
  import.meta type-checks (the package stays CJS-published; tsconfig only drives
  lint:types, esbuild owns the build).
- `start` now runs the built dist (node dist/index.js) since src is no longer
  directly node-runnable as ESM.

Also fix runtime-common/ignore.ts to import the `ignore` CJS package via a
namespace (`.default ?? ns`) instead of a default import: esbuild bundling
downstream consumers (vscode-boxel-tools) couldn't synthesize the default.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…r boot

The ts-node->node codemod's entry capture grabbed the closing `)` of a
`$(ts-node … secret)` command substitution and appended `.ts` after it:

  MATRIX_REGISTRATION_SHARED_SECRET=$(node "…/matrix-registration-secret.ts").ts

so the secret value was suffixed with a literal `.ts`. The realm-server then
failed to come up, never bound :4201, and every integration suite that needs the
dev-services stack (Host / Realm-Server / Matrix / Live Tests and downstream)
cascaded. Verified locally: with the secret restored, the realm-server starts
under native node and serves :4201 (base/skills modules + indexing).

Fix the three mise service tasks (realm-server, test-realms, realm-server-base)
and harden the codemod's entry pattern to exclude shell metacharacters
( ) ; & | ) so a command substitution can't be mis-captured again.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Three native-ESM gaps the CI integration stack surfaced, each cascading widely:

- handlers/handle-indexing-dashboard.ts used `require.resolve('morphdom/…')` at
  module load. `require.resolve` (no `require(` token, no `require.main`) slipped
  past the earlier sweeps; it threw `require is not defined` and crashed the
  worker-manager on import -> no indexing -> realm-server never reached :4201
  readiness. Resolve via createRequire.
- matrix synapse start (support/synapse/index.ts, support/docker.ts) and
  realm-server utils/jwt.ts imported CJS packages as `import * as fse/jsonwebtoken`
  and then read members off the namespace. Under native ESM those members live on
  the CJS default, so `fse.stat`/`sign`/`verify` were `undefined` — synapse failed
  to start (`fse.stat is not a function`) and JWT signing/verification would no-op.
  Switch to default imports.

Verified each member resolves under native node; matrix and realm-server
type-check clean.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

The migrations dir now carries a package.json pinning it to
type:commonjs (so the CJS migration files keep working under the
package's new type:module). node-pg-migrate's CLI default ignore
pattern only skips dotfiles, so `pnpm migrate` loaded that package.json
as a migration named "package" with no up/down function and threw
"Unknown value for direction: up".

Pass the same --ignore-pattern the programmatic runner in pg-adapter.ts
already uses, so the CLI path (DB setup, CI migrate steps) skips
package.json and .eslintrc.js too.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

The migrations dir carries a package.json pinning it to type:commonjs
(so the CJS migration files keep working under the package's new
type:module). node-pg-migrate's CLI default ignore pattern only skips
dotfiles, so the test-pg seed builder loaded that package.json as a
migration named "package" with no up/down function and threw "Unknown
value for direction: up", failing realm-server DB setup in CI.

4ae9312 fixed the `pnpm migrate` package.json script the same way, but
the create_seeded_db.sh seed builder has its own inlined node-pg-migrate
invocation that was missed. Pass the same --ignore-pattern so the seed
path skips package.json and .eslintrc.js too.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

`import * as QUnit from 'qunit'` yields an ESM namespace whose `.assert`
is undefined under native Node ESM (qunit is CJS; the real object is the
default export). ts-node's esModuleInterop previously masked this. The
helper's `QUnit.assert.codeEqual = codeEqual` then threw "Cannot set
properties of undefined (setting 'codeEqual')" at import time, aborting
the entire realm-server test suite.

Switch to the default import (`import QUnit from 'qunit'`) — the form
every realm-server test file already uses — so `.assert` resolves.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

`import QUnit, { module, test } from 'qunit'` fails under native Node
ESM with "The requested module 'qunit' does not provide an export named
'module'" — qunit is CJS and exposes module/test as properties of the
default export, not as ESM named exports. Because tests/index.ts
requires every test file at load time (TEST_MODULES only filters which
run), this one file aborted all six realm-server shards.

Destructure module/test from the default import, matching the 142 other
realm-server test files that already use `import QUnit from 'qunit'`.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

The realm-watch-stop integration fixture spawned a child that called
`require('ts-node').register(...)` to load watch-process-registry.ts
from source. ts-node was removed in the ESM migration, so the child
threw "Cannot find module 'ts-node'", exited 1, and both
realm-watch-stop tests failed with "fake watcher exited early".

Node 24 strips TypeScript types natively, so the registry .ts loads via
a plain require with no loader. Verified the fixture boots
(FAKE_WATCHER_READY, registers, stays alive) and all three tests pass.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

ts-node was dropped from every package's deps, but invocations, a broken
process-kill pattern, and stale comments/docs lingered. Sweep them:

Functional
- boxel-cli `bin/boxel.js`: the dev fallback `require('ts-node')` is gone;
  the CLI ships as a built CJS bundle (source uses CommonJS-only globals
  like __dirname and isn't native-ESM-runnable), so error clearly when
  dist is missing instead of throwing "Cannot find module 'ts-node'".
- `mise-tasks/lib/dev-common.sh`: the orphan-sweep matched
  `node_modules.*--transpileOnly (worker|main|prerender)`, which never
  matches the native `node main.ts` / `node worker-manager.ts` /
  `node prerender/*-server.ts` processes — so `mise run kill-all` and dev
  cleanup leaked the realm/worker/prerender ports. Match the native entries
  (by name + a distinctive flag) instead.
- `packages/postgres/Dockerfile`: ts-node → node for fix-migration-names.ts.
- `.github/workflows/boxel-cli-publish.yml`: ts-node → node (3 sites).
- delete `packages/realm-server/scripts/run-test-modules.cjs` — an unused
  qunit runner that still bootstrapped ts-node (the suite runs via
  run-qunit-with-test-pg.sh / `node tests/index.ts`).

Docs / comments / log strings
- start-*.sh: "about to exec ts-node" log lines now say node; header
  comments drop the ts-node tsconfig/binary rationale.
- mise service comments, worker-manager.ts, opencode.ts, find-package-root.ts,
  fixtures.ts, context-search + canusetool usage strings, boxel-cli + ai-bot
  READMEs, and the indexing-diagnostics skill: ts-node → node.

Left intentionally: the esm-codemod tooling, accurate "without ts-node"
notes, pg-adapter's historical esModuleInterop rationale, and the
phase-1-plan historical design doc.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Follow-up to the ts-node cleanup: the last references were explanatory
comments/docs, not usage. Reword them to describe the native-Node
behavior directly instead of contrasting with ts-node.

- .eslintrc.js / erasable-syntax-selectors.cjs: describe the
  "erasable TS for --experimental-strip-types" rule without naming ts-node.
- pg-adapter.ts: the node-pg-migrate `.default` unwrap is still required
  (native ESM binds the CJS namespace object; the runner is on `.default`)
  — keep the cast, but explain it purely in native-ESM terms.
- fake-watcher fixture + software-factory phase-1 plan: native `node`
  wording, no ts-node.

Only scripts/esm-codemod/** still mentions ts-node, by design (it's the
migration tooling).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…t tests)

The two `/_post-deployment` maintenance-endpoints tests stubbed
`compareCurrentBoxelUIChecksum` / `writeCurrentBoxelUIChecksum` on the
`import * as` namespace of boxel-ui-change-checker. Under native ESM that
namespace is read-only, so sinon threw "ES Modules cannot be stubbed" and
both tests failed (the only red in realm-server shard 4).

Group the two functions on an exported mutable object `boxelUIChecker` and
have the post-deployment handler call through it; the tests stub the
object's methods. This matches the codebase's stubbing convention (stub
methods on objects/instances — global.fetch, stripe.subscriptions — never
module namespaces) and preserves the handler's behavior. Verified the stub
intercepts calls through the object under native node.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Env-mode CI (host shards, Live Tests, Software Factory) drives the
prerender through puppeteer, which in env-mode relies on puppeteer's
bundled Chrome (env-vars.sh only points PUPPETEER_EXECUTABLE_PATH at a
system Chrome in non-env-mode). actions/cache caches the pnpm store but
not ~/.cache/puppeteer, and on a store cache hit puppeteer's postinstall
skips the Chrome download — so the prerender threw "Could not find Chrome
(148.0.7778.97)", the standby pool stayed empty, base indexing never
finished, and /base/_readiness-check hung (host shards timed out at
000000). Main only passed because its warm cache still had Chrome.

- init: explicitly `puppeteer browsers install chrome` after install
  (idempotent; no-op when already cached) so every job that runs the
  prerender has Chrome regardless of cache state.
- Software Factory: `playwright install --with-deps` so the apt system
  libraries are present too ("Host system is missing dependencies to run
  browsers").

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

`pnpm --filter <pkg> exec puppeteer browsers install chrome` went through
pnpm's recursive-exec path and dropped the `chrome` arg in CI (puppeteer
printed its usage and exited 1), which failed the init step for every
workflow. Use a non-recursive `pnpm exec` from the package directory
instead — the same form SF's `pnpm exec playwright install` uses, which
forwards args correctly.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

puppeteer's postinstall leaves an empty browser folder
(~/.cache/puppeteer/chrome/linux-<ver>) with the download skipped, so the
prerender fails "Could not find Chrome" and a plain `browsers install`
refuses with "folder exists but executable is missing — All providers
failed". Remove any partial install first, then download cleanly (the CDN
works — main's postinstall downloads the same version fine).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Abandon the bundled-Chrome approach for env-mode CI — puppeteer's
postinstall leaves an empty cache folder and an explicit `browsers
install` wouldn't reliably land Chrome where the prerender looks, so the
env-mode prerender kept failing "Could not find Chrome".

env-vars.sh already points PUPPETEER_EXECUTABLE_PATH at the system Chrome
in standard mode; the block was just inside the non-env-mode guard. Move
it to the "regardless of env-mode" section so env-mode uses the runner's
/usr/bin/google-chrome too (it exists — the non-env-mode jobs already use
it). Gated on the binary existing and the path being unset, so it's a
no-op-or-correct in env-mode prod (the prerender container installs
google-chrome-stable, and an explicit override still wins). Revert the
init browser-install step it replaces.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…l module

@cardstack/boxel-cli/api mapped to raw api.ts source. That resolves under
some loaders (vitest, plain node via the workspace symlink realpath) but not
Playwright's ESM worker loader, which failed loading software-factory's
.spec.ts files with "does not provide an export named 'BoxelCLIClient'",
taking down every SF Playwright shard.

Bundle api.ts to dist/api.js (esbuild CJS — named exports stay detectable by
Node's cjs-module-lexer, so ESM consumers keep `import { BoxelCLIClient }`)
and point exports["./api"] at it; the types condition still resolves to
api.ts, so type resolution is unchanged and no .d.ts is needed. Add a
`build:api` script and run it before software-factory's node + playwright
tests so the built module is present (only SF consumes this subpath).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

`pnpm --filter @cardstack/boxel-cli build:api` runs only buildAPI() to
rebuild the API surface quickly for cross-package test consumers
(software-factory). But the bundled dist/api.js imports content-tag
transitively, and content-tag reads its wasm with
`${import.meta.dirname}/content_tag_bg.wasm` — i.e. from boxel-cli's
dist/. The wasm copy lived inside buildCLI() and only ran on a full
build, so software-factory's `test:node` / `test:playwright` saw
ENOENT on dist/content_tag_bg.wasm.

Hoist the copy into a standalone step that always runs.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Same root cause as eb42f81 / 40ad5cf — qunit is CJS and exposes
module/test as properties of the default export, not as ESM named
exports. Under native Node ESM, `import { module, test } from 'qunit'`
throws "does not provide an export named 'module'". Because
tests/index.ts requires every test file at load time (TEST_MODULES
filters which run), this one file aborted all six realm-server shards.

Match the convention used by the other realm-server test files.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Native Node ESM doesn't expose `__filename`. The sibling realm-server
tests use `basename(import.meta.filename)` for the same module label;
match that convention.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

After rebasing onto main (which bumped puppeteer to 25.0.2 and removed
ts-node from realm-test-harness), `pnpm install` strips the residual
ts-node entries that lingered in pnpm-lock.yaml from the pre-rebase
state. No package.json changes — just lockfile alignment.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>