/**
 * sanitizeDraft — minimal pre-submit cleanup for chat-composer drafts.
 *
 * Mirrors the conservative behaviour ChatGPT / Claude / Telegram
 * actually ship: clean the obvious junk the user didn't intend to
 * send, touch nothing that *could* be intentional.
 *
 * **What we DO touch:**
 *
 *   1. Trim leading/trailing whitespace (spaces, tabs, newlines,
 *      NBSP). The user typing `\n\n   hello   \n` meant `hello`.
 *   2. Normalise line endings — `\r\n` / `\r` → `\n`. Pasted Windows
 *      / old-mac text gets the same internal shape, so the LLM
 *      tokeniser and markdown renderer see one canonical form.
 *   3. Strip zero-width / invisible characters that web-paste
 *      smuggles in: ZWSP (U+200B), ZWNJ (U+200C), ZWJ (U+200D),
 *      BOM / ZWNBSP (U+FEFF). They're invisible, break LLM
 *      tokenisation, and the user never meant to type them.
 *
 * **What we DO NOT touch (and why):**
 *
 *   - **Internal whitespace runs** (3+ spaces, tabs, blank lines).
 *     Code indentation depends on these. ChatGPT preserves them as
 *     typed — "  if (x):\n    return" stays four-space-indented.
 *     Collapsing them is the path to subtly broken code snippets.
 *
 *   - **Bidi override marks** (U+200E LRM, U+200F RLM, U+202A..U+202E).
 *     Legitimately used in Arabic / Hebrew / mixed-direction text.
 *     Stripping silently breaks RTL users. If a specific deployment
 *     wants to block them as a security measure, do it at that layer
 *     with explicit user-visible feedback.
 *
 *   - **Tabs vs spaces** beyond rule 1. Could be either code or
 *     prose; without parsing markdown we can't tell.
 *
 *   - **Emoji, mentions, URLs, code spans** — passthrough text.
 *
 * The function is intentionally tiny — every rule earns its keep
 * with a concrete "user pasted X from Y, got nonsense" story.
 *
 * **Idempotent**: `sanitizeDraft(sanitizeDraft(x)) === sanitizeDraft(x)`.
 */
export function sanitizeDraft(input: string): string {
  if (!input) return '';

  // Strip zero-width invisibles. Done FIRST so the trim below sees
  // the real content edges — a stray ZWSP at the start would
  // otherwise count as non-whitespace and survive the trim.
  // U+200B ZWSP, U+200C ZWNJ, U+200D ZWJ, U+FEFF BOM/ZWNBSP.
  let s = input.replace(/[​‌‍]/g, '');

  // Normalise line endings.
  s = s.replace(/\r\n?/g, '\n');

  // Trim outer whitespace (includes \n, \t, NBSP via String.trim).
  s = s.trim();

  return s;
}

/**
 * Convenience predicate: true when the draft is non-empty AFTER
 * sanitation. Use to gate Send buttons / Enter submits so an empty
 * or whitespace-only draft never produces a real message.
 *
 * Cheaper than sanitizeDraft(input).length > 0 only marginally —
 * we still allocate the cleaned string. Kept as a named helper for
 * call-site clarity.
 */
export function isSubmittableDraft(input: string): boolean {
  return sanitizeDraft(input).length > 0;
}