chore(dev): add opt-in Prettier pre-commit hook and editor format-on-save

[pjdoland]

Summary

Contributor PRs land red on CI fairly often, and the failure is almost always the same shape: a single Markdown, TypeScript, or CSS file that was edited without running jlpm lint first. jlpm lint:check gates Prettier across the whole tree (**/*{.ts,.tsx,.js,.jsx,.css,.json,.md}), so one unformatted file is enough to turn the run red even when the actual change is fine.

This adds two small, independent, opt-in layers that move that formatting fix from CI back to the moment of editing or committing. Neither changes what CI enforces; they just catch the drift earlier. jlpm lint remains the manual catch-all.

The two layers are split into two commits so either can be adopted or dropped on its own:

• 9efdcfe Editor format-on-save (zero-risk, no dependencies). .editorconfig sets baseline whitespace rules (and mirrors the existing package.json 4-space override). .vscode/settings.json enables format-on-save with Prettier, scoped per language so Python and any formatter a contributor already uses are left untouched. .vscode/extensions.json recommends the matching extensions. Only those two .vscode files are un-ignored in .gitignore; all other per-user .vscode files stay ignored.
• 17d5256 Pre-commit hook. A postinstall script runs Husky, which registers a pre-commit hook that invokes lint-staged. lint-staged runs prettier --write --ignore-unknown on staged files only and re-stages the result, so the committed version is already formatted.

Solution

The load-bearing decisions:

• postinstall, not prepare. Husky's documented setup uses prepare, but I confirmed empirically that Yarn Berry (this repo's jlpm) does not run the root workspace's prepare on install, while it does run postinstall. So prepare would silently never install the hook here.
• husky || exit 0 guard. Husky is a devDependency, so anyone installing the published @plmbr/notebook-intelligence npm package does not receive it; a bare postinstall: husky would fail their install with "command not found." The || exit 0 keeps it a no-op in that case, and also where there is no .git (the sdist to wheel build), so it never affects packaging. exit 0 is portable across the POSIX shell, Yarn Berry's portable shell, and cmd.exe.
• Prettier-only hook. lint-staged runs only Prettier (safe to auto-apply and idempotent). ESLint and Stylelint rule violations are deliberately left for the contributor and CI to surface, since --fix for those can make non-trivial changes. Because eslint-plugin-prettier and stylelint-prettier both enforce the same Prettier config, formatting fixed by the hook satisfies the Prettier dimension of all three linters.
• "*": "prettier --write --ignore-unknown". Keyed on * rather than a duplicated extension glob, so it cannot drift from prettier:base. --ignore-unknown skips file types Prettier has no parser for, and Prettier still honors .prettierignore, so staged build output or binaries are left untouched.

Testing

• Verified the hook end-to-end on real commits: a deliberately misformatted Markdown file is reformatted and re-staged before the commit completes; a binary/unknown file staged alongside it is left byte-for-byte unchanged. Both commits in this PR were themselves made through the hook.
• Confirmed husky || exit 0 exits 0 when the husky binary is absent (consumer case) and when there is no .git (sdist build case).
• Confirmed the .gitignore negation tracks exactly .vscode/settings.json and .vscode/extensions.json while keeping other .vscode files ignored (git check-ignore).
• jlpm lint:check, jlpm tsc --noEmit, and jlpm jest (328 tests) all green. No Python changed, so pytest is not applicable.

Risks / follow-ups

• Both layers are opt-in by design. The hook can be skipped per commit with git commit --no-verify or disabled entirely with HUSKY=0 jlpm install.
• The hook formats any Prettier-parseable staged file, which is a superset of the types CI checks. This is harmless (formatting is idempotent and CI only checks the narrower set), but it means a staged .yaml or .scss, for example, would also be formatted.
• Optional, not included to avoid touching packaging config: the newly tracked dotfiles ride along in the source sdist (not the wheel, and not the npm tarball). They could be added to the [tool.hatch.build.targets.sdist] exclude list, but the sdist already carries other dev files (CONTRIBUTING.md, lint configs), so pruning only these would be inconsistent. Happy to add it if you would prefer a leaner archive.
• yarn.lock grows by the husky and lint-staged dependency trees; the change is purely additive (generated with jlpm).