feat: make the browser build tree-shakeable, drop UMD

[digitalSloth]

Summary

• Drop the pre-bundled UMD/webpack browser build (bundle.browser.js/.mjs) and ship modular ESM only. Bundlers (Vite, Rollup, Webpack) now resolve dist/index.js directly, enabling tree-shaking — a build importing only Zenon/Address/Hash no longer pulls in KeyStore/KeyFile/argon2.
• Replace Node's crypto module and ed25519-hd-key with @noble/hashes (hand-rolled SLIP-0010 derivation), and the crypto/crypto-browserify AES-GCM driver in Encryptor with @noble/ciphers. These were the last Node-builtin dependencies blocking a browser bundler from resolving the modular SDK at all.
• KeyFile's Argon2 KDF keeps its Node/browser split: native argon2 for Node/CLI (unchanged, ~21ms), argon2-browser's bundled subpath for the browser (avoids the wasm require() that broke esbuild's dependency optimizer).
• package.json's exports["."] simplified to {types, default}; sideEffects: false added.

Fixes #33.

Breaking changes

• No UMD window.ZnnSDK build. Script-tag consumers need a bundler or an ESM CDN.
• require("znn-typescript-sdk") needs Node ≥ 20.19 (unflagged require(esm)), or switch to import.
• KeyFile's browser Argon2 loads as a dynamic chunk — bundlers serving the SDK need output.publicPath (or the Vite equivalent) set correctly.

Next release should be 2.0.0.

Verification

• npm test — 650 passing, no fixed-vector changes (existing keyStore/keyFile vectors prove old mnemonics/keyfiles still work).
• deriveKey independently checked against the official SLIP-0010 ed25519 test vectors (including the 2147483647' boundary); non-hardened paths rejected, index overflow fails closed.
• Encryptor's AES-256-GCM cross-tested against Node's own crypto output over 500 randomized rounds — byte-identical ciphertexts, bad-tag decryption still throws before returning plaintext.
• npm run lint / npm run build clean.
• Tree-shaking confirmed with a scratch webpack build: main chunk dropped from ~1.3 MB to ~196 KiB with KeyFile/argon2/KeyStore strings absent when unused.
• esbuild --platform=browser resolution confirmed clean (no wasm/fs/stream errors).

Docs

READMEs and AGENTS.md updated: PoW file paths repointed to lib/pow.{js,wasm} (previously dist/browser/), UMD references removed, browser-builds section rewritten.

[0x3639]

Tested this against zwap.fun (the wallet-less consumer from #33: Vite 8.1.4 / Rolldown, production minify) plus a Node and a browser KeyFile smoke. It closes #33 for us. Details below, one suggestion at the end.

zwap.fun build, before vs after

Built from a tarball of this branch (npm pack → npm i ../znn-typescript-sdk-1.0.5.tgz → vite build), same machine, same config.

	1.0.5 (published)	this branch
SDK-bearing chunks	bundle.browser 944 KB + argon2 634 KB	sdk-node 105 KB + htlc 70 KB
argon2 chunk	always emitted	none
direct eval (vm-browserify)	present, Rolldown [EVAL] warning	none
elliptic / asn1 / KeyFile / KeyStore / Encryptor strings in output	present	none
total JS in dist/assets	2.06 MB	676 KB
build result	ok (but blob)	ok — the [REQUIRE_TLA] failure on argon2-browser/dist/argon2.wasm that the modular build used to hit is gone

The only bundler noise left is two warnings from dist/pow/pow.js (see suggestion). zwap.fun can also shrink its vite-plugin-node-polyfills include from ["crypto","buffer","stream","util"] to ["buffer"] and still build (total JS 608 KB).

KeyFile in a real browser

Minimal Vite 8 app, no polyfill plugin, importing KeyStore/KeyFile/Zenon/Address, run in Chrome against both vite build && vite preview and the dev server:

• argon2-bundled.min is emitted as a separate lazy chunk (46 KB) — only because this app reaches KeyFile; zwap.fun's build has no such chunk.
• KeyFile.setPassword(p).encrypt(ks) → .decrypt(json) round-trips the mnemonic (encrypt ~170 ms, decrypt ~140 ms).
• Wrong password throws (aes/gcm: invalid ghash tag) before returning plaintext.
• The UMD default interop works as written under Rolldown in both prod and dev.

Node, from the tarball (Node 22.12)

• import("znn-typescript-sdk") and require("znn-typescript-sdk") (require-esm) both load, same Zenon instance.
• KeyStore.fromMnemonic, KeyFile encrypt/decrypt via the native argon2 variable-specifier import (47 ms), PoW init() + generate() from lib/ all work.

Crypto replacements, cross-checked independently

• AES-256-GCM: 200 randomized rounds (random key, nonce, 1–96 byte plaintext) — Encryptor.encrypt output is byte-identical to Node aes-256-gcm with AAD "zenon", and ciphertext produced by either side decrypts with the other. Existing keyfiles are safe.
• SLIP-0010: 100 random hardened paths, depth 1–5, indices up to 2³¹−1 — Crypto.deriveKey and Crypto.getPublicKey match ed25519-hd-key@1.3.0 exactly.
• npm test: 650 passing; lint clean.

Code read-through: ensureSha512Sync() is called on every @noble/ed25519 entry point that needs it (getPublicKey, sign; there is no verify in Crypto), so sideEffects: false is safe. No other module-level side effects in src/. dist/wallet/keyFile.d.ts doesn't reference the argon2-bundled ambient module, so consumers' typecheck is unaffected by src/types/*.d.ts not being emitted.

Suggestion (non-blocking): silence the Node-only imports in pow.ts

Rolldown still resolves the fs/url/path dynamic imports under isNode():

[plugin rolldown:vite-resolve] Module "path" has been externalized for browser compatibility, imported by ".../znn-typescript-sdk/dist/pow/pow.js"
[plugin rolldown:vite-resolve] Module "fs" has been externalized for browser compatibility, imported by ".../znn-typescript-sdk/dist/pow/pow.js"

and, when a polyfill plugin is present, emits a never-fetched ~51 KB url polyfill chunk. Adding the node: prefix and /* @vite-ignore */ removes both warnings and the chunk (48 fewer modules transformed) with no behaviour change — PoW init()/generate() re-verified on Node 22 from the packed tarball:

const { readFileSync } = await import(/* webpackIgnore: true */ /* @vite-ignore */ "node:fs");
const { fileURLToPath } = await import(/* webpackIgnore: true */ /* @vite-ignore */ "node:url");
const { dirname, join } = await import(/* webpackIgnore: true */ /* @vite-ignore */ "node:path");

Also checked esbuild 0.28 --platform=browser --bundle on dist/index.js: clean exit with both the current and the prefixed form. Ready to cherry-pick from 0x3639/znn-typescript-sdk@ee56e57 (branch feat/tree-shakeable-browser, based on this PR's head).

Small behaviour changes worth a line in the 2.0.0 notes (none affect KeyStore/KeyFile)

• Crypto.randomBytes(n) now sits on crypto.getRandomValues, which caps a single call at 65,536 bytes (QuotaExceededError above that, Node and browsers alike). Node's randomBytes had no such cap. Internal callers only ask for 12–64 bytes; a chunked loop would restore the old range if you want to keep the public API identical.
• The ed.etc.sha512Sync hook is now installed lazily inside getPublicKey/sign instead of at import time. A consumer that imports the SDK and then calls @noble/ed25519's sync API directly (e.g. ed.verify) before constructing a KeyPair will hit etc.sha512Sync not set, where 1.0.5 happened to have set it for them. Reasonable trade for sideEffects: false; a Crypto.verify wrapper would give those users a supported path.
• hexToBytes in deriveKey is stricter than the old Buffer.from(seed, "hex") (throws on 0x-prefixed or odd-length hex instead of silently truncating). Only invalid input is affected.

Pre-existing, not introduced here

• In an environment that is both isBrowser() and Node (jsdom under vitest/jest, Electron renderer with nodeIntegration), the Emscripten glue in argon2-bundled.min.js keys off process.versions.node, takes its Node path, and aborts trying to fetch //argon2.wasm — the process exits rather than the promise rejecting. The old argon2-browser entry behaves the same, and the PR's parity test works around it by nulling process.versions.node. Worth a README note for people running KeyFile under jsdom; zwap.fun's vitest suite doesn't touch KeyFile so it's not affected.

Nits

• tsconfig.json ts-node.files: true: the test script runs via tsx, so this is inert; harmless either way.
• README: under Vite the Argon2 chunk URL follows base automatically, so the output.publicPath note mainly applies to webpack users.

Happy to run a canary of the 2.0.0 tarball against zwap.fun again before release.