feat(studio): the release date is a month and a day (SCIENCE-1 answered)

[aliakherati]

SCIENCE-1 is answered, and the date follows from the answer

Zonal-mean, monthly climatology (your call, 2026-08-18), selected by the month of the release. So
the reduced product will be (lat × month × level) — order 86k values per field, a few MB, small
enough to commit with a checksum rather than fetch at run time. Recorded on #53 and in
OPEN_QUESTIONS.md; task 1.1 is unblocked.

That decides the shape of the date, which is what this PR builds.

	before	after
entered	day_of_year: int (1–366)	month + day_of_month
derived	—	day_of_year, on a fixed non-leap calendar
year	n/a	none — a monthly climatology is an average over years

A monthly climatology has no year, so accepting one would imply the run used that year's weather.
And day_of_year must be derived rather than entered alongside the month: entering both permits a
config whose month and day disagree — the climatology read at one date and the sun at another.

Non-leap is a declaration, not an oversight

studio/science/calendar.py owns it. The same date in a leap year is day 173, and the solar
declination differs by ~0.01° — far below the uncertainty in anything it feeds. The fixed calendar
buys a config that means one thing forever, and the test pins the value that ties it to the archive:

21 June  →  day 172        # exactly what the paper ensemble uses

29 and 31 February are refused, not clamped — in the schema as well as in the derivation, so the
error names the two fields the user typed rather than a calendar function.

Verified in the browser

month  6 day 21 -> day 172        month  2 day 30 -> 422 "February has 28 days on the
month 12 day 21 -> day 355                            non-leap calendar (SCIENCE-1)"
month  1 day  1 -> day 1

And the SZA panel follows the date, with the physics checking out at 30°N:

June      13.74 h daylight   min SZA  6.55°     (|30 − 23.44| =  6.56)
December   9.93 h daylight   min SZA 53.42°     ( 30 + 23.44  = 53.44)

Also

The model seam now refuses an unresolved day_of_year the same way it refuses an unresolved
concentration — a raw config reaching the model would otherwise pick a solar declination of its own.

346 Python Tier-A (+19), 55 vitest, lint and mypy clean. No UI code: the date fields render on
stage 1 with their units and provenance because the form is generated.

The hash moved again inside the still-unmerged 0.3.0 (…6d3a73 → …7813ae). Same run — the same 21 June
— described differently.

Stacked on #92, which is stacked on #90. Merge in that order and I'll rebase anything GitHub
flags.