fix(dockerfile): pin Bun to 1.3.13-slim to dodge 1.3.14 ARM64 crash

[devxoul]

Background

A bug report from an Apple-silicon (ARM64) host: the container reports a Bun
version, but every attempt to run an external package crashes.

$ bun run dist/src/platforms/webex/cli.js --help
[1]    3 Aborted                 bun run dist/src/platforms/webex/cli.js --help
EXIT: 134

bunx, bun add, and bun run <pkg> all abort with SIGABRT (exit 134) or a
NotDir internal error; even bun upgrade self-fails with StackOverflow. The
agent-messenger CLIs (agent-webex, etc.) were effectively unrunnable inside the
container.

Root cause is upstream Bun, surfaced by our own tag policy. The base image
FROM used the floating oven/bun:1-slim tag, which silently rolled the
base onto Bun 1.3.14. That release has two ARM64 Linux regressions that hit
exactly the external-package launch path:

• A self-referencing symlink created in createFakeTemporaryNodeExecutable
  (/tmp/bun-node-*/node -> bun -> bun -> ...) -> ELOOP / SIGABRT
  (oven-sh/bun#30711).
• require()-of-ESM aborting with a PAC IB trap (silent SIGTRAP/SIGABRT)
  (oven-sh/bun#30281).

The fix for #30711 (oven-sh/bun#30713)
is not yet in any stable release, so there is no forward fix to roll to -- the only
remedy today is to stop landing on 1.3.14.

Summary

Pin the base image to an exact patch, oven/bun:1.3.13-slim -- the last
known-good release -- instead of the floating 1-slim tag. Floating tags make
base-image rebuilds non-deterministic; a single upstream patch release can break
every installed agent on the next rebuild, which is exactly what happened here.

The pin lives in one BUN_IMAGE constant that feeds both the prebuilt base
image (buildBaseDockerfile) and the inline dev/test path (renderInlineHead),
so production and dev mode can't drift onto different Bun versions.

The three CI Bun pins (ci.yml, release.yml, base-image.yml) move
1.3.14 -> 1.3.13 so the runner validating/releasing the image uses the same Bun
as the container. They were already pinned (not latest); this just holds them
at the known-good patch and documents why.

What this does NOT do

• Does not rebuild or publish a new typeclaw-base image. That happens via
  the Release workflow; this PR only changes what the next release will build.
• Does not touch the affected agent folders on hosts -- they pick up the fix
  on their next typeclaw start once a release ships the new base.
• Does not work around the upstream bug in our own code; it avoids the broken
  Bun version entirely.

Review notes

• Load-bearing change: src/init/dockerfile.ts BUN_IMAGE = oven/bun:1.3.13-slim.
  Invariant to verify -- bun run scripts/emit-base-dockerfile.ts | grep ^FROM
  now prints FROM oven/bun:1.3.13-slim.
• The drift guard in dockerfile.test.ts asserts base and per-agent Dockerfiles
  share the same FROM; tests updated to the new tag. The two
  not.toContain('FROM oven/bun:...') assertions were widened to FROM oven/bun:
  so they stay meaningful regardless of the pinned patch.
• The two start.test.ts "stale Dockerfile" input fixtures still use 1-slim on
  purpose -- they represent an old on-disk Dockerfile that refreshDockerfile
  overwrites.
• Bump to the next stable once install: stop nested --bun from self-referencing its node shim oven-sh/bun#30713 lands; the BUN_IMAGE comment
  flags that follow-up.

[typeey]

The pin is applied through the shared Dockerfile constant that feeds both inline and base-image generation, and the updated assertions cover the important drift cases. The matching workflow Bun pins and the existence of the oven/bun:1.3.13-slim amd64/arm64 image make this safe to land from a code-review perspective.

[devxoul]

Closing: this fix is incorrect. Reproduced the crash on ARM64/OrbStack and proved via strace that it is NOT a Bun version regression — Bun 1.3.13 crashes identically. The real root cause is the bwrap sandbox's --tmpfs /proc: the spawned child of bunx reads /proc/self/fd and /proc/self/maps which don't exist under the empty tmpfs proc, surfaced as Bun's NotDir. Fix moves to src/sandbox (opt-in real-proc strategy). Superseded by a follow-up PR.

[typeey]

Got it — thanks for the repro and strace. That makes sense; the Bun pin isn’t the root cause here. I’ll treat this as superseded by the sandbox-side fix in src/sandbox.