refactor(repo): drop bun for vitest, node, and esbuild

[haydenshively]

Part 2 of 2 for CRTR-2822. Stacked on #129 — review that first; this PR's base is
chore/migrate-to-pnpm, so the diff shown here is only the bun removal.

Completes the toolchain consolidation: 136 test files move to vitest, bots bundle with esbuild and
run on Node, images carry no bun binary, CI drops setup-bun.

What moved

Surface	Replacement
141 bun:test files	vitest; mock()→vi.fn(), spyOn→vi.spyOn, mock.restore()→vi.restoreAllMocks(), .toInclude→.toContain
soltag bunfig preloads	soltag/vite in tests, an esbuild onLoad plugin in each bot's scripts/build.ts
bun src/index.ts	node --env-file-if-exists=.env dist/src/index.js
Bun.env (59)	process.env
Bun.spawn (10)	node:child_process, with the deterministic SIGKILL + exit-promise semantics preserved
Bun.serve (3)	node:http, behind an adapter that keeps the Web-standard (Request) => Response handlers
Bun.file, Bun.sleep, Bun.which, Bun.argv, import.meta.dir, import.meta.main	readFile, node:timers/promises, resolved binaries, process.argv, fileURLToPath, an argv[1] comparison
bun $ shell (3 deploy scripts)	execa; the secret still goes via stdin ($({ input })), never argv
global confirm()	node:readline/promises; a non-TTY declines rather than hanging CI
Bun.build (@repo/contracts)	esbuild, same soltag plugin

Four latent bugs this surfaced

1. 24 floating expect(...).rejects assertions. bun's typings hid them, so they never asserted
   anything. oxlint's no-floating-promises catches them once vitest typings apply.
2. Anvil fork ports collided. bun's runner was serial, so fixed ports were only deconflicted
   within a bot — blue and midnight-liquidation both wanted 8545, midnight-queue and
   market-making both wanted 8546. vitest runs files in parallel, so every fork suite now claims a
   distinct port, with the registry documented in the harnesses.
3. fetch.preconnect in http-json.utils.ts was a bun-only extension to fetch in production
   code.
4. tryCatch mishandled every execa call (found in review). It branched on
   fn instanceof Promise, and execa's subprocess is a thenable that is not a native Promise while
   its typings declare it extends Promise<Result> — so the check type-checked and then took the
   sync path, calling the subprocess object. assertCli() reported "Railway CLI not found" for a
   railway that exists, before Railway was ever invoked. It now branches on callable and assimilates
   thenables via Promise.resolve; the ad-hoc Promise.resolve(...) wrappers that a few call sites
   already carried are removed, so all of them read the same.

prestart stays, for the local path only

start runs dist/src/index.js, which a clean checkout does not have. The images build at
image-build time and their CMD never fires a pre-script, so each bot keeps
prestart: pnpm --filter "{.}..." --if-present run build purely for the documented
pnpm --filter <bot> run start path. {.}... scopes it to that bot plus its workspace dependencies,
so one bot's start does not build the other three.

Two bounds deliberately raised

A tsx cold start costs ~1.3s against bun's ~0.1s. The market-making subprocess tests get a 30s
ceiling, and the FIFO fail-closed probe a 10s bound (a FIFO read blocks forever, so any finite
completion still proves the loader fails closed). Both still prove what they were written to prove;
the interpreter genuinely got slower, so the ceilings moved rather than the assertions.

Evidence

• Same test files run. Diffed vitest's collected file list against every tracked *.test.ts:
  141 vs 141, zero delta — including scripts/check-jsdoc.test.ts, the one test outside a
  test/ tree, which vitest's default glob would silently have dropped.
• Non-vacuity, all 11 projects. Added a failing probe to one file per project and confirmed each
  project actually executed and failed it; all files restored, grep clean.
• @repo/contracts output byte-identical across Bun.build→esbuild: abis/Midnight.json,
  abis/Morpho.json, dist/index.d.ts all match sha256. The solc pin from chore(repo): migrate from bun to pnpm #129 is what makes this
  hold.
• Bundles actually run on Node. All four reach their own fail-loud config check, and the soltag
  bots' bundles contain zero residual sol` templates with 32 inlined bytecode blobs each.
• Fork suites genuinely run. With RPC_URL_8453 supplied, the midnight fork suites spawn anvil,
  deploy the soltag-compiled Executor and run a real liquidation — end-to-end proof of the
  soltag/vite transform. loadEnv was verified not to clobber an already-exported value (so CI
  cannot silently skip) and to omit the key rather than set an empty string.
• Secret path probed at runtime: secret appears in argv? false / secret arrived on stdin? true.
• The tryCatch fix is pinned by tests that fail without it. Two new cases in
  packages/utils/test/helpers/tryCatch.test.ts reproduce fn is not a function against the old
  implementation; a real execa call through the real tryCatch then returns error: null for a
  present binary and still errors for a missing one.
• prestart verified from wiped dists for all four bots — each builds its workspace dependencies
  and its own bundle, then reaches its own fail-loud config check.
• Full suite green: 1459 pass, 1 skip, 141 files, fork suites included.
  pnpm lint 0/0 · pnpm knip clean · pnpm -r run typecheck 12/12 · jsdoc:build exits 0.

main is merged in

main gained #121/#124/#125 after this branch's merge-base. Because a stacked PR's mergeability
is computed against main rather than the immediate parent, that left GitHub reporting the PR as
conflicting and suppressing its checks entirely. Merging main in fixes both and lands the anticipated
follow-up early: 8 more test files ported off bun:test, 2 more where main added mock( sites to
files already converted, Bun.file/Bun.serve/Bun.sleep replaced in the setup-state test, and
3 more floating expect(...).rejects awaited — the same latent bug class, still hidden by bun's
typings on main.

Docker is verified

All three images build, run as uid=1000(node), carry no bun binary (node v24.14.1,
pnpm 11.1.1 only), and reach their own fail-loud config check from node dist/src/index.js with the
bundles built at image-build time. The non-root switch itself comes from #129.

Not verified

Nothing outstanding.

Follow-ups

• Of the six open PRs that add new bun:test files and Bun.* calls, docs(market-making): separate architecture documentation #121, feat(market-making): support router ratifiers #124 and fix(market-making): batch invalidation with Midnight multicall #125 have
  merged into main and are already absorbed here (see "main is merged in" above). test(market-making): add MKT-1491 workflow e2e coverage #112, fix(bot-kit): stamp pending txs with the broadcast-time head #116 and
  feat(market-making): add docker image, compose, and docker hub publish #123 still need the mechanical fix on rebase — feat(market-making): add docker image, compose, and docker hub publish #123 is the sharpest, adding a bun-based
  market-making/Dockerfile plus a Docker Hub publish workflow that must be ported to node-slim.
• bots/market-making/src/infrastructure/invalidation/batch-offer-invalidation.utils.ts (from fix(market-making): batch invalidation with Midnight multicall #125,
  via main) is in neither check-jsdoc.ts nor typedoc.json. Pre-existing on main, not introduced
  here, so it is left alone — but it is the same registration gap the review caught on
  bundle-failed.error.ts.
• Bump oxlint with oxlint-tsgolint >= 0.24 (pin comment left in the catalog).

🤖 Generated with Claude Code

[devin-ai-integration]

Devin Review found 3 potential issues.

View 3 additional findings in Devin Review.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: ce08fe4a6e

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[haydenshively]

Heads-up: GitHub reports this stacked PR as conflicting, and git disagrees

GitHub shows mergeable: CONFLICTING / mergeStateStatus: DIRTY, which is also why no Checks run has fired on this branch since 13:27Z — the workflow only triggers on pull_request opened/synchronize, and GitHub skips that when it cannot compute the merge.

There was a real conflict, and it is now fixed. main gained #121/#124/#125 after this branch's merge-base, and a stacked PR's mergeability is computed against main, not against the immediate parent — which is why #129 stayed green while this one went dirty. main is merged in as of ddacbad, along with the mechanical port of what it brought (8 more files off bun:test, Bun.file/Bun.serve/Bun.sleep, and 3 more floating expect(...).rejects).

After that merge, git reports no conflict from any angle:

conflicts vs main:                 0
behind main by:                    0
base (chore/migrate-to-pnpm) is an ancestor of head:  yes
compare base...head:               status "ahead", behind_by 0

GitHub has not recomputed. The usual escape hatches are both closed for a stack:

• PUT /pulls/130/update-branch → 403 Updating a stacked PR's branch via this endpoint is not supported.
• gh pr edit --base … → Cannot change the base branch because the pull request is part of a stack.

A rebase onto the updated parent (verified byte-identical tree) and a fresh push did not clear it either.

What is needed: someone with the PR open in the browser hitting the stack's Restack / Update control, which is the only path the API leaves for a stacked PR. Merging #129 first would also collapse the stack and moot it. Once the state clears, Checks will run on the next push.

Everything is verified locally in the meantime: 1459 pass / 1 skip / 141 files (fork suites included), vitest collects exactly the 141 tracked *.test.ts with zero delta, pnpm lint 0/0, knip clean, 12/12 typecheck, jsdoc:build exits 0, and all three images build, run as uid=1000(node), and reach their fail-loud config check.

[haydenshively]

Resolved — the stack is rebased and checks are running

My previous comment said this needed the Restack control in the browser. That was wrong: the fix is gh stack (the official github/gh-stack extension), which the repo's stack was created with.

gh stack checkout 131   # stack #131: main <- #129 <- #130 <- #134
gh stack rebase
gh stack push

All three PRs now report mergeable: true (blocked is just the pending required review), and Checks fired on all three for the first time since 13:27Z.

Two notes on what the rebase needed by hand:

• Linearizing dropped the merge commit, and with it the bun:test → vitest ports for the seven files main had added. They are restored as their own commit (refactor(market-making): port main's new test files off bun:test), and the resulting tree is byte-identical to the merge-based state that was verified.
• The rebase tried to replay main's own 14 commits on top of #129 — they were only reachable through that merge commit and are already in the new base's ancestry, so they were dropped from the todo rather than re-applied.

Verified after the restack, per branch:

	#129	#130	#134
own contribution vs. pre-rebase	identical	tree byte-identical	identical (same 25 files)
tests	—	1459 pass / 1 skip / 141 files	1522 pass / 1 skip / 142 files
lint · knip · typecheck	✓	0/0 · clean · 12/12	0/0 · clean · 12/12