feat: an alerts workload in the Weather starter, a range outside the key

[krzysztof-smartdataengines]

What

sde-weather run --workload alerts and sde-weather-ts run --workload alerts: in every iteration,
besides the write and the point read every workload makes, the run reads its station's readings at
or above 95% humidity - a bounded page, a count and an exact celsius summary - and compares each
with the generator exactly.

Why

Every read the starter could drive filtered on key columns. mixed, point and analytics fix
the station, which the declared key (station, then at) already serves; fleet ranges over at
across stations, which a key reordered to (at, station) serves. Nothing asked a question only an
added index can answer, so a control plane measuring the starter's traffic had no evidence for the
one physical change that is built in place (docs/in-place-index.md) - and a demonstration could
not show one.

alerts is that question: a range on humidity, a field outside the key, within one station. The
window says so in its per-shape section - filtered_on: [{"equal": ["station"], "range": "humidity", ...}] for the range reads and the aggregates - names only, no values.

Exactness

Humidity is 30 + sequence mod 70, so a reading alerts when sequence mod 70 >= 65: five in every
seventy, known from the sequence number alone. alert_expected / alertExpected computes the first
page (one station, so key order is sequence order), the count and the exact decimal total. Before
the first alert the page is empty and the summary's total is null, as docs/logical-reads.md
specifies for a summary of no values on every engine - the live runs cover that iteration too.

Tests

• Brute force in both languages: the expectation against every generated row filtered by the
  threshold, for four (rows, page limit) pairs including none and exactly one alert.
• The run loop in both languages against an in-memory session that answers reads as an engine does
  (equality, range, key order): a faithful engine completes; a page that ignores the range, a count
  that ignores it, a total off by a cent, and a zero total for no values each refuse the run with
  the check that caught it.
• Live, both engines, both languages (test_demo_starter_live.py, weather.live.test.ts): two
  iterations of 40 rows - none alert in the first, five in the second - complete, and the window's
  filtered_on is exactly the station equality and the humidity range.
• The CLI offers every workload the runtime runs (WORKLOADS / workloads, one list each).

Mutations (both languages, harness in the private acceptance record): 14 applied, all killed; both
controls (comments) survived.

make check with both live engines: see the checks below.

🤖 Generated with Claude Code