Hand back a fax as samples, not as the bytes it was compressed into

[tannevaled]

reader v0.4.2 → v0.5.0. No code change.

reader v0.5.0 decodes /CCITTFaxDecode — Group 3 one-dimensional, Group 3
mixed, Group 4. Images gets its data from reader.Decode, so a scanned
picture used to come back with Filter set to the fax name and Data holding
bytes still compressed. Image.Filter is documented as "empty for data this has
unfiltered into plain samples"; for a fax that is now true.

Measured

Two binaries calling extract.Images on every page, each verified with
go version -m before running:

old  github.com/go-pdfkit/reader v0.4.2   github.com/go-pdfkit/pdffont v0.3.0
new  github.com/go-pdfkit/reader v0.5.0   github.com/go-pdfkit/pdffont v0.3.0

Corpus: all 2 268 forms in /Users/Shared/pdfforms. Prevalence below is over
the 1 633 real documents; the 635 vendor test fixtures are counted separately.

	files
compared	2 268
inventory differs	74
… of the 1 633 real documents	71 (4.3%)
… of the 635 vendor fixtures	3
changed which images it finds	0
open/read status changed	0

By issuer, of the 71 real: fr-cerfa 26, us-opm 23, fr-impots 19,
us-uscourts 2, uk-govuk 1.

2 861 images changed, and the filter name they arrived under says why:

images	Filter before	after
2 592	/CCF	(empty — samples)
269	/CCITTFaxDecode	(empty — samples)

Their Data goes from 527 375 bytes still compressed to 12 708 706 bytes of
samples (24.1×). One image, verbatim from the inventory of
fr-cerfa/cerfa_10134.pdf:

before:  filter=CCITTFaxDecode  bytes=372
after :  filter=              bytes=1892

Nothing else moved. No DCTDecode or JPXDecode image changed — a JPEG is
still handed back as a JPEG, which is what the package promises.

The ImageFilter contract change

reader.ImageFilter no longer answers true for /CCITTFaxDecode or /CCF. I
checked all five consumers of this fleet: nothing calls ImageFilter. The
only two places that switch on a filter name are extract's own
noteImage, which just records whatever Decode reports, and latex's
picture(), which tests for DCTDecode and JPXDecode only and is unaffected.

Not measured

Wall clock and peak memory. Three other corpus jobs were running on this machine,
so any timing figure would be fiction. The output diff above is unaffected by
load.

Gates

go vet, gofmt, exact 100% statement coverage, and all nine cross-compile
targets pass locally with GOWORK=off CGO_ENABLED=0.