Store a star's distance error in two bytes, so the card prints the error its catalogue published

a81dd49 kept the square root of the relative error in 255ths. A step was a few per cent of the
error itself, which moved the last digit the card prints: over the cached Gaia answers, 2 822 of
the 53 209 stars whose card prints an error printed another one than their published parallax
error gives (Gaia DR3 1415230383034813824 "112 +- 2 pc" for 3), and Rigel, 3.78 +- 0.34 mas in van
Leeuwen 2007, read "265 +- 23 pc" for 23.8.

The column is now a Uint16 in 65 535ths, placed before the photometry byte so its view stays
two-byte aligned whatever the star count; BYTES_PER_STAR_META goes from 16 to 17, the build.ts
round trip tolerance to half a 65 535th, and the README names the column. The same count gives 12,
each on a rounding half. In the running app Rigel reads "265 +- 24 pc".

The cost: stars-meta.bin goes from 7 288 512 to 7 744 044 bytes, and gzip -9 from 3 117 639 to
3 623 379, half a megabyte more to download, since the extra byte is noise-like. One reviewer
judged the one-byte figure within the errors' own accuracy; this takes the other two's view that
the card should print the published number.

Control: 255 steps again fails "keeps an error close enough that the card prints the published
one" alone.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

README.md

@@ -209,9 +209,9 @@ The star catalogue ships as two binary column stores plus a small JSON file, not
 objects. At 68 388 stars the old encoding — one JSON object per star, its eight key names
 repeated each time — would have been about 17 MB to download and parse before the first frame.
 Splitting it puts the numbers in `stars.bin` (positions, handed to the GPU verbatim) and
-`stars-meta.bin` (id, magnitude, colour index, spectral-type index, the band and colour system
-those were measured in or whether the colour was read off a temperature, and the distance's
-relative error), and leaves `stars-index.json`
+`stars-meta.bin` (id, magnitude, colour index, spectral-type index, the distance's relative
+error in two bytes, and the band and colour system those were measured in or whether the colour
+was read off a temperature), and leaves `stars-index.json`
 holding only the strings, with the ~2 600 distinct spectral classifications collapsed into a
 dictionary. The result is 2.6 MB for 7.8× the stars. `star-catalog.ts` defines the layout once
 and both the ETL and the app use it, so the writer and the reader cannot drift apart.