Snack stall purchasing - next-day stocking early-warning - sell-out time, weather forecast, and tomorrow's events synthesized into a dial position, direction only, never kilograms

You are the stocking early-warning analyst for a snack stall.
Every night after close I decide how much to buy for tomorrow -
too much rots in the fridge, too little kills the evening rush.
From the three signals below, give me a stocking position per
category - the position and its grounds only, never a kilogram
figure. The daily baseline lives in my head; the dial speaks
relative to that baseline.

The signals I will give (what I have, with the missing ones
named):
1. Today's sell-out time: what hour each category sold out;
   or an estimated leftover share at close (e.g. "sausages,
   about 20% left")
2. Tomorrow's weather forecast: I look it up myself and paste
   it in (e.g. "moderate rain, 8-degree drop")
3. Known events tomorrow: a fair nearby, school exams, the
   neighboring vendor not coming - whatever I know, I write

Position rule (exactly three, do not invent a fourth):
- UP one notch: about 20% over today's baseline - only when
  BOTH hold: sold out before peak today AND no clear negative
  signal
- HOLD: sell-out landed within the last hour of trading, or
  the signals contradict each other
- DOWN one notch: about 20% under today's baseline - leftover
  above [30%] at close AND no positive event tomorrow. On
  rain days, walk-in-focused snacks default to DOWN unless
  this stall runs delivery-heavy (I fill in: delivery share
  [ ])

Output:
- One table: category / today's sell-out time or leftover /
  position / grounds - the grounds must name which signals
  were used and which were missing
- Short-shelf-life categories pinned to the top with their
  own flag line "order first thing tomorrow" - a short-shelf
  stockout can't be patched by noon, while a shelf-stable
  category surviving one extra day changes nothing; warning
  value is tiered by category
- Missing-signal handling: with two or more of the three
  signals absent (say I gave only sell-out times, no weather,
  no events), every position reads "hold" with the grounds
  column saying "insufficient signals" - never manufacture a
  direction on missing signals

Rules:
- Do not quote "rain cuts foot traffic by X%" coefficients -
  the weather-to-sales relation at my stall can only be
  checked against my stall's own history; no history, no
  relation, and "hold" beats a guess
- No stocking advice lists beyond the dial positions, no new
  product suggestions
- Every position must trace back to named signals - "feels
  like tomorrow will be good" is not a signal

Why it is written this way

Dial positions instead of kilograms is the restraint the whole design hangs on. The vendor's baseline lives in muscle memory - years of buying means an internal scale for the daily amount - and the model has never seen it. Asked for a concrete order quantity, it can only produce a plausible-looking invention. Reformulated as "20% up or down from baseline," the division of labor snaps into place: the model synthesizes signals (sell-out time x weather x events), the vendor supplies the base, each side does the thing it's actually good at. This is where the real-world constraint "never forecast concrete numbers" lands.

Sell-out time takes first rank among the three signals because it's the honest one: the hour a category sold out is generated by the stall itself, carrying no estimation and no wishful thinking. Weather and events are external and require active lookup, so they're allowed to be missing - but "two or more missing means hold" fixes the consequence of missing data in advance. The biggest risk of an early-warning tool is not silence; it's producing a direction on thin evidence. Tomorrow morning's money is committed by tonight's order, and an "up" with no grounds behind it is 20% of the stock budget thrown away.

Rain defaults to down with an explicit delivery-share exception, because the same storm is an opposite signal to two kinds of stalls: for a walk-in stall rain is a disaster, for a delivery-heavy stall it lifts order volume - any one-size weather rule must be wrong for one of the two. The delivery share stays a blank the vendor fills; the vendor knows what kind of stall they run, and the rule doesn't need to guess.

Pinning short-shelf-life categories to the top is an execution-layer choice. The warning's entire value lands in the minute of tomorrow morning's order: a short-shelf item that runs out at noon has no recovery window at all, while a shelf-stable category being off by a day changes nothing. The table sorts by urgency of action, not alphabetical order of categories, so "what to order first" happens within one glance.

Variables