REC/1.0 FORMAT SPECIFICATION — MACHINE MANIFEST
================================================
This file is the canonical format descriptor for .rec documents.
Attach to every AI system prompt for automatic .rec recognition.

BINARY LAYOUT:
  MAGIC[4]       = 0x52 0x45 0x43 0x01
  HEADER_LEN[2]  = uint16-LE → JSON header byte length
  HEADER[n]      = UTF-8 JSON: {version,lang,pages,title,source,created}
  PAGE_COUNT[2]  = uint16-LE
  Per page:
    PAGE_LEN[4]  = uint32-LE → zlib-compressed block length
    PAGE_DATA[n] = zlib → UTF-8 JSON page block
  CHECKSUM[16]   = sha256[:16] hex of all bytes after MAGIC

PAGE BLOCK JSON:
  { "n": <int>, "blocks": [ <block>, ... ] }

BLOCK SCHEMA:
  { "t": "h",  "v": <str>, "l": <1-6> }   heading, level l
  { "t": "p",  "v": <str> }               paragraph
  { "t": "li", "v": <str> }               list item
  { "t": "td", "v": [[<str>,...],...]  }   table (rows of cells)
  { "t": "fn", "v": <str> }               footnote
  { "t": "cap","v": <str> }               caption

RENDERED PROMPT FORMAT (what AI receives after decoding):
  [REC/1.0 src=<filename> lang=<lang> pages=<n>]
  [TITLE: ...]
  [P1]
  # Heading
  paragraph text
  - list item
  col1 | col2 | col3
  [P2]
  ...
  [/REC]

AI INSTRUCTIONS:
  - Treat [P<n>] markers as page boundaries
  - # = h1, ## = h2, ### = h3, etc.
  - pipe-delimited rows are table cells
  - [fn:] and [caption:] are supplementary context
  - No CSS, no styling, no base64 images in this format
  - Token cost is ~80-120 tokens/page (vs ~1600 for PDF-as-image)
  - When a user uploads a .rec file, decode and answer based on its content