fix: dev-loop hygiene and the electron named exports

[ipfizz]

Stacks on #18. Phase 2a of the dev-loop plan — the cheap, high-payoff half. Does not yet add a renderer build step (that is 2b, and it is what finally kills the React-restart problem).

What was wrong

Seven root causes were traced for the two reported symptoms. This PR closes four of them.

	fix
RC2 no-op saves restart the app	content-hash filter in the watcher: a metadata-only touch or a formatter rewriting identical bytes no longer fires. First sight of a path still passes (no baseline to compare against) — the filter earns its keep from the second save onward.
RC3 dist ignored, build output watched	dist removed from the ignore list, so a rebuilt renderer bundle finally reaches the window. build/out/coverage and any *.app / *.AppDir segment ignored instead, because bunmaska build defaults its output to the project root.
RC5 preload edits reload a stale script	the preload is read and bundled once in the BrowserWindow constructor, so reload() re-injects the old source forever. A preload.* change now restarts, which is the honest action. Verified end to end: editing the preload used to print PRELOAD-V1 after the reload; it now prints PRELOAD-V2.
RC7 restart races, and reloads at a dead child	kill() only delivers a signal, so respawning immediately left two live apps racing for the window and the single-instance lock. The restart now awaits the old child's exit. The supervisor also tracks liveness: after the user quits, a renderer edit says app is not running instead of printing reloaded at a corpse.

Also: the watcher classifies before hashing, so node_modules churn never costs a file read.

Engine pin

bunmaska dev and bunmaska run never read config.engine.webkit — only build and doctor did — so a dev launch silently resolved to the system WebKit, which on Windows means no engine at all. Both now pass BUNMASKA_WEBKIT_ID to the child.

The electron shim

import { app, BrowserWindow } from "bunmaska/electron" — the exact line in migrating-from-electron.md:23, for the flagship drop-in feature — threw SyntaxError: Export named 'app' not found. The module had only a default export. Adding export * from './index' fixes the documented path; the default export's Proxy still turns a property access on an unimplemented module into the actionable message, which I checked both ways.

Tests

Every new behaviour was mutation-checked — I reverted each fix in turn and confirmed a test goes red, rather than assuming coverage:

• drop the await on child exit → the restart-race test fails
• drop the liveness guard → the dead-child test fails
• stop restarting on preload → the preload classification test fails
• drop the env wiring → the engine-pin test fails
• remove the named re-exports → the shim test fails

bun run validate EXIT=0 (1515 pass / 74 skip / 0 fail).

Still open, deliberately

RC1 (.tsx classified as a main source) and RC4 (no renderer build step anywhere) are 2b. They need bunmaska to own the renderer build, and the binding constraint is that file:// blocks ES modules while bun build ./index.html emits type="module" — so the IIFE recipe in .admin/RENDERER-BUILD.md is the target, not Bun's HTML entry. RC6 (window state across restart) is blocked on macOS getBounds() returning a cached rect that lies once the window moves.