windows: how to actually schedule the publish, not just what to copy

[mspinola]

The repo shipped run-publish.cmd and push-panel.cmd on 2026-08-04 with a README line saying to chain them behind errorlevel guards after the cotdata producer task, and nothing saying how. cotdata has docs/WINDOWS_SCHEDULING.md; this had no equivalent, which a setup walkthrough hit immediately.

Adds docs/WINDOWS_SCHEDULING.md and links it from the README's scheduling section.

Deliberately thin. It defers to cotdata's doc for the schtasks syntax, the Friday release trigger, restart-on-failure and the troubleshooting list. A second copy of a living document drifts from the first and nobody notices until someone diffs them, which this repo has already paid for once (CLAUDE.md, "The copy lost 104 lines and nobody noticed for a day").

Four things that were written down nowhere

• Neither crowdmon step gets its own timer. --require-final defers until the Norgate Data Updater has pulled the Finals, and restart-on-failure retries 6x10min, so a 20:55 prices task can finish an hour later. A publish on a 21:30 timer builds last week's panel and reports success doing it.
• The call trap. cmd treats a bare .cmd invocation from inside a .cmd as a transfer of control, not a subroutine, so without call the wrapper runs run-prices.cmd alone, the other three steps never execute, and the task exits 0. No error, nothing in any log.
• Verify from manifest.json, not Last Run Result, which is not a success signal here (a deferred price run exits non-zero by design). provenance.built_at is the field that catches a dead schedule: COT is weekly, so a stopped job produces no new report week to notice, and a panel can be current on the week and months old on the clock.
• Register by repointing the existing prices task, so the restart-on-failure setting already configured covers the whole chain. That re-runs the price fetch on a retry, which is idempotent, so it is correct rather than merely tolerable. Said plainly rather than presented as free.

It also places the cotdata store push between publish and panel push, so the server never briefly holds a panel derived from data it has not received.

Verification

Docs only, no src/ or tests/ change. tests/test_references.py passes against the worktree, so the §D14 citation resolves. The four suite totals in CLAUDE.md are unmoved: no test was added or removed, and no parametrised guard grew its corpus.

🤖 Generated with Claude Code