Keep the time shape of a column placed after another

[jarichb]

Problem

A date/timestamp column with a dependency (an explicit after, or a created/updated/closed name pair) lost its time shape: order_row_times replaced its drawn value with latest_required + expovariate(mean 3 days), so business_hours on the column (or --business-hours) had no effect. A gap past the window was clamped to as-of midnight, piling rows onto one timestamp.

Repro (5,000 rows, seed 1, as-of 2026-10-01; scheduled_for has after: booked_at + business_hours: true):

	before	after
scheduled_for in weekday 08:00–17:59	28.9% (uniform)	71.8% (shaped_alone: 73.2%)
with --business-hours	~29%	72.7%
rows at exactly 2026-10-01 00:00:00	57	0
rows before booked_at	0	0
median gap	~2 days	~2 days

Fix

• New place_after in generate/timeline.py: the gap still decides roughly when. With business hours (the column's own hint or the run's), weekends are kept 3 in 10 and a timestamp's hour is drawn from HOUR_WEIGHTS on the gap's day. A draw past the window is drawn again; after 20 tries the value is placed evenly between the column it follows and the window end.
• order_row_times takes the run's time_profile. when columns filled on day 0 (_moment, settle) use the same helper.
• A dependent column still takes growth and seasonality from the column it follows.

Determinism

The README's promise is now per version: same model, seed, --as-of and options give the same bytes on any machine and any day with the same model2data version. A release that changes what a seed produces says so in the changelog. tests/test_determinism.py still fails on any unintended byte change; its docstring and message say when re-recording is allowed.

Only one recorded case changed (skew: saas_platform with --skew 0.8), because the cases run at 40 rows. Real models of a few hundred rows or more with a timestamp chain will change, and the CHANGELOG says so.

Tests

• New tests in tests/test_timeline.py: business hours kept under after (column hint and run-level), no pile-up at as-of midnight, place_after bounds.
• Full suite: 1,841 passed. ruff clean; ty diagnostics unchanged at 34.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!