`indd::convert` now returns `Conversion::previews`: InDesign's screen
preview (chunk 0x170D -> class 0x1708 -> chunk 0x119 -> raw data) of each
placed graphic without `Contents`, one entry per preview file with the
`Self` of every graphic that uses it. IDML has no place for a preview, so
the package is unchanged; an application can show the preview when it
cannot read a linked file. `convert_into` and `indd audit` do not read
previews.
Output: `diff_outputs.py --exclude own/` over 23 files (the fixtures and
part of the open corpus): 0 differing outputs, 0 changed warnings or
failures.
Paragraph and character styles get Imported (the u16 after the style
kind) and StyleUniqueId (the GUID after the name) from the INDD, and the
values every non-root style has in IDML, such as PreviewColor, from
observation. The values generator leaves root styles out of the style
paths.
Value coverage on trustworthy pairs: 84.07 % to 84.29 %.
indd audit converts a document while recording which objects, chunks and attribute IDs the converter reads, then lists what it did not read: classes, chunks, byte-stream objects, attribute IDs per kind of list, codes of known attributes without an IDML value, and strand run kinds, with counts and the conversion's warnings. --tsv gives rows for scripts. tools/audit_corpus.py audits every distinct corpus file and ranks the items by files affected. Without the audit, recording does nothing and conversion output is unchanged.
- InDesign 5.0: read document chunk 0x13501 with the 3.0/4.0 layout
(four UID lists); the 6.0 layout failed on 17 files.
- Reject files that are shorter than their database pages as truncated,
and check the signature before the size so that short files of other
kinds are reported as not INDD files. The corpus tests accept both.
- Hyperlink page destinations: write view setting 0 as Fixed (567 of 567
in the pairs); leave out a zoom outside the schema's 5-4000 %.
- Leave out, with a warning, values the IDML schema does not allow: ink
neutral density -1, writing script 0xFFFFFFFF, negative drop shadow
size and other out-of-range transparency values. The IDML writer now
returns its own warnings.
- Bullet list: the font style can be followed by 1 and four bytes.
- InDesign 2.0: in-object strings start with 1, not 2.
- InDesign 2.0-5.0: XML tag names are a flag byte and a string, in
either byte order.
- InDesign 3.0 and 4.0: the document's hyperlink lists start at offset
0 and include page destinations.
- XMP: the packet can be shorter than its object; take its length.
- InDesign 1.x files (no signature, DOCUMENT at 0x5C) are reported as
not supported; files whose page directory lies beyond the database
pages as having no object database.
- Schema: leave out hyperlink sources that hold an anchored object, and
the characters U+FFFE and U+FFFF, which XML does not allow.
tests/corpus.rs accepts the two rejections above and skips the paths
listed in the local corpus/exclude.txt.
Remove the sample documents, PDF and licence copies from tests/fixtures/.
tests/fixtures/manifest.json lists each file's download URL (pinned to a
commit for GitHub sources), size and SHA-256. tools/fetch_fixtures.py
downloads them into the git-ignored tests/fixtures/files/ and verifies
them. The fixture tests read from there and skip when it is absent.
compare.py --all now also converts the INDD and INDT files that have no
same-version IDML (any version, either byte order), validates them and
counts converter warnings by kind. validate.sh runs Jing once per schema
instead of once per file, which makes validating the whole corpus take
about 3 minutes.
Two schema failures showed up in files without an IDML:
- Gradient swatch midpoints are stored as positions from the start of
the gradient. IDML wants them relative to the gap between two stops,
as for opacity gradient stops. Values outside 13-87 are left out with
a warning.
- The style kind is a u16, not a u32. One InDesign 7.5 template has a
non-zero u16 after it, which made 151 character styles paragraph
styles.
The byte order flag applies to object data only: database pages stay
little-endian. Object data is decoded in the file's byte order, and in
big-endian data the string tag and the flag byte before it are swapped.
The two big-endian samples are also the only files from before
InDesign 7.0, and some of their layouts differ: page transform and
bounds chunks, font records, XML tag names and nodes, story strands and
40-byte guide records are read; text frame settings, document
preferences, a composite font and a section style are left out with a
warning. The 40-byte guides of an InDesign 7.0 file are now read too.
docs/format/big-endian.md records the evidence. The corpus tests now
cover all byte orders.
Build a document model (layers, spreads, master spreads, pages, page
items with path geometry, text frame threading, stories with paragraph
and character style runs, style names and hierarchy) and write it as an
IDML package with a minimal stored-ZIP writer.
Add tools/compare.py, which converts every corpus pair and compares the
result with the reference IDML. First run: all 351 little-endian files
convert; 921 of 1120 stories match exactly.
Decode the database pages: page trailers and types, the logical page
table, the object B+ tree keyed by (UID, segment), data pages, slotted
pages and continued records. Database::object returns any object's byte
stream. Add `indd objects` and `indd object` commands.
Every object of all 351 little-endian corpus files reads without an
inconsistency. Big-endian databases report Unsupported for now.
Add a project CLAUDE.md with the clean-room rules.
Select the active master page by sequence number, read the database page
count, iterate the contiguous objects after the database pages, and
extract the XMP packet. `indd info` now prints these fields.
The layout comes from Adobe's BSD-licensed XMP Toolkit SDK and is checked
against all 352 corpus files and the fixtures. Add Adobe's BlueSquare.indd
sample (BSD) as a fixture and record the XMP Toolkit as a permitted
reference.
Decode the signature, byte-order flag and InDesign version from the INDD
header, add an `indd info` command, and add smoke tests on four
open-licensed sample files plus an optional local corpus test.
Document the clean-room method and the header layout with the evidence
for each field.