fix(agent): hash code assets without Web Crypto so AI tools work over non-secure origins

[cakelesscoder]

Summary

The runtime code-asset tools hash content with crypto.subtle.digest. crypto.subtle only exists in a secure context — HTTPS or localhost.

Instatic is self-hosted, so reaching the admin over a LAN or Tailscale IP on plain http:// is an ordinary deployment rather than an edge case. In that context crypto.subtle is undefined, so the first code-asset tool call throws crypto.subtle is undefined and the agent stops mid-turn.

This adds a dependency-free sha256Hex in src/core/utils/sha256.ts and uses it for the read→patch optimistic-concurrency fingerprint in codeAssetTools.ts.

That fingerprint is self-consistent and client-only — it answers "did this file change between read and patch" and never bears cryptographic security — so a pure-JS SHA-256 is a drop-in, and it returns byte-identical digests in every context. No new dependency; it's ~100 lines of standard SHA-256.

Verification

• bun run build
• bun test
• bun run lint
• Docker/deployment check — not relevant (no deployment surface touched)

New test src/__tests__/utils/sha256.test.ts checks the implementation against native WebCrypto across the empty string and known vectors, multi-byte UTF-8, every message-length residue around the 55/56/64-byte padding boundaries, and single-character sensitivity.

Checklist

• Tests cover behavior changes.
• Docs were updated when behavior, config, deployment, or public surfaces changed. — no doc-visible surface changed; the fix is internal to an existing tool.
• No compatibility shim was added for old pre-release behavior.
• No secrets, local databases, uploads, or generated artifacts are included.

---

Split out of a larger branch so it can be reviewed on its own — it's independent of any AI provider and fixes the existing agent tooling on non-secure origins.