build(publish): gate npm publish on the suite, and rebuild dist at pack time

[andrei-hasna]

Closes todos 4efcbd8a. Two lines in package.json, no source change.

npm publish shipped whatever dist/ happened to be on the publishing machine. Nothing forced a rebuild and nothing ran the tests, so a stale, partial or locally-patched dist/ would reach the registry as-is and the tarball would not be derivable from any commit. On a package whose entire purpose is redacting credentials, the security claim then rests on the publisher checking by hand.

I paid that cost twice tonight publishing 0.1.5 — typecheck and the suite run manually, on the exact published tree rather than on my branch. The next publisher would have neither the gate nor a reason to suspect one was missing.

The change

hook	command	hazard it closes
prepublishOnly	bun run typecheck && bun run test	publishing a red tree
prepack	bun run build	shipping a dist/ that is not a clean rebuild from source

build already opens with rm -rf dist, so prepack makes the shipped artefact a function of the commit rather than of the machine. It also covers npm pack, so a tarball someone inspects is built the same way as one that ships.

Verified in both directions

A gate that cannot block is the thing this change exists to remove, not to add — so both states were exercised.

POSITIVE — dist/ deleted on purpose, then npm publish --dry-run:

• prepublishOnly ran (typecheck and suite output present in the log)
• prepack rebuilt dist from source
• pack contained all 24 files including dist/index.js
• the rc=1 there is cannot publish over the previously published versions: 0.1.5 — the registry refusing a duplicate version, not the gate

NEGATIVE — scratch copy at unpublished version 99.99.99, so the gate is the only possible failure source, with one deliberately false assertion planted:

• npm publish --dry-run exits 1 at prepublishOnly
• the planted failure is what stops it
• total files appears zero times — refused before a tarball is built

Gates on this tree: typecheck rc=0, bun test rc=0, build rc=0.

Sibling survey — the task asked for this rather than fixing one package in isolation

Probed 60 of the 165 first-party packages in the bunfig excludes inventory, reading each package's published scripts from the registry rather than a local checkout:

prepublishOnly present   36     <- positive control: the probe finds the hook where it exists
prepublishOnly ABSENT    16
not published             8
no build-at-pack hook    37     <- tai's exact hazard

The 36 are what make "absent" an observation rather than a broken read.

Scope, not completeness. The survey is alphabetically bounded and stopped at 60 — every one of the 16 falls between access and entities, so this is a sample and the true count is higher. Filed separately with the names; not fixed here, because sixteen packages' publish behaviour does not belong in a tai PR.

Refs: todos 4efcbd8a

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}