format: text and Markdown files can be written in UTF-16

[donislawdev]

txt and md gain two settings, shared in internal/format/textenc: encoding
(utf-8, utf-16le, utf-16be) and bom. Both default to what these formats
already wrote, so a recipe that says nothing gets the same bytes it got before -
pinned by five new golden values, with every one of the 57 existing hashes
untouched.

tfg generate --format txt --size 4kb --set encoding=utf-16le --set bom=true

An odd size is refused, not rounded

Every character takes two bytes in UTF-16, so only an even number of bytes can be
a whole file. Measured on three readers, in three languages, on the same file cut
to an odd length:

reader	whole file	cut by one byte
Python 3.14.7, strict decode	ok	rejected
Node v26.8.1, TextDecoder(fatal:true)	ok	rejected
.NET UnicodeEncoding(throwOnInvalidBytes)	ok	rejected
all three, lenient	no replacement characters	one, silently
Get-Content	2 lines	2 lines, looks identical

The last row is the one that decided it: the cut file looks exactly like the
whole one. So half of all sizes become unreachable and the refusal names the
size either side of the one asked for. The registry mechanism for that already
existed - SmallestAccepted walks to a fixed point and already handled PNG
refusing a band above a size it accepts.

Single byte encodings are deliberately absent

The filler vocabulary is ASCII, so a file written as latin-1 is byte for byte the
same file as UTF-8 - compared with cmp. That would be a setting that changes
nothing. They belong with content that is not English, which is a separate change
with its own golden bytes.

Two things this turned up that were not the job

• TestEveryFormatSurvivesItsReferenceTool skipped the structural check for any
  format declaring no reference tool, so the new checkers would have been dead on
  arrival - with the guard green and reporting both formats as covered. It was
  invisible while the only formats without a reader were also the only ones
  without a checker.
• A strict decode alone blesses a file with its byte order reversed, because
  0x74 0x00 read the other way is a valid CJK character. What catches it is
  asking what the characters are, and only txt was asking.

txt and md also gain the structural check they never had. The comment saying
they had nothing to check against stopped being true rather than being wrong, and
was rewritten with the reason rather than deleted. The checker is told the
encoding rather than sniffing it, for the reason the CSV dialect is told.

Verification

• whole suite green, gofmt, vet, preflight --quick all 12 checks
• nine mutations, all caught
• 285 of 285 checks on the built binary, 41 of them new
• 48 generated files decoded strictly by an independent decoder

🤖 Generated with Claude Code