[ColinToft]

Why

Switching Colin's milk food's only measure (`g`) from `Atomic` to `Weight` dropped its diary entries to 1/100 of their grams. This PR records the server's actual rule so nobody repeats that, and makes the recipe-ingredient refusal warn against it.

The server's rule (verified live on ZZ foods, 2027-03-30, Oct 13 2026)

A diary entry is stored as a count of its measure. Its grams are:

• `Weight` measure: count x measure value
• `Atomic` measure: count x 100, whatever its value

Logging always sends grams, and the count is derived with the same rule, so fresh entries are always right. Measured with the server's day totals (400 kcal/100 g food):

food	measure	logged	counted
g-only Atomic	g	50 / 150 g	200 / 600 kcal
g-only Weight	g	50 / 150 g	200 / 600 kcal
`create_custom_food` shape (Weight scoop 30 g + server g)	scoop	30 / 90 g	120 / 360 kcal
same	g	50 / 150 g	200 / 600 kcal
Atomic scoop 30 g	scoop	30 / 90 g	120 / 360 kcal

Editing a measure re-reads existing entries:

• Atomic -> Weight (value 1): 180 g -> 1.8 g; back to Atomic: 180 g again
• Atomic scoop (30) -> Weight: 90 g -> 27 g
• Weight scoop (30) -> Atomic: 90 g -> 300 g; Weight g -> Atomic: 50 g -> 5000 g
• resizing a Weight measure 30 -> 60: 90 g -> 180 g; resizing an Atomic one: unchanged
• replacing the measure list with a new Weight g: the server re-points the entries to it and they read 1.8 g

So PR #11's foods (Weight serving + server `g`) are diary-correct; no code fix is needed. The client never changes a measure's type.

Changes

• Comment at `_WEIGHED_MEASURE` with the rule and the numbers.
• `update_custom_food` docstring: only Weight resizes rescale; types are re-sent unchanged.
• `_check_recipe_ingredient`'s refusal now says not to switch the type, and to make a new food instead. Test asserts it.
• README: the same warning.

Tests

`uv run pytest -q`: 830 passed. ruff check/format clean.

🤖 Generated with Claude Code