windows: the two scheduler templates that actually get the panel to the server

[mspinola]

Adds the two Windows scheduler templates that actually get the damage panel onto the dash server. Requires mspinola/cot-analyzer#16, which deprecates the Mac-side script this was previously (and uselessly) wired into.

Why this exists

The optional fourth payload added to cot-analyzer/scripts/push_data_cache_to_server.sh in #14 was dead code. That script pushes from the Mac, and cotdata/docs/SYNCING.md documents the Mac as a read-only replica: one Windows server produces everything and feeds two replicas, the Mac over SMB and the dash server over SSH. So the panel had no path to production at all.

CROWDMON_STORE is a separate store root from COTDATA_STORE, so cotdata's own push-to-server.cmd does not carry it and no change to its exclusion list would. Hence a second push script rather than a flag.

The two templates

Following cotdata's convention exactly — copy out of the repo into a scheduler directory, fill the REPLACE_WITH_* markers there so a git pull cannot clobber them, wire to Task Scheduler, chain behind errorlevel guards.

template	does
docs/examples/windows/run-publish.cmd	builds the panel via the portable bin/publish_damage.py
docs/examples/windows/push-panel.cmd	rsyncs CROWDMON_STORE to the dash server over SSH

Order: prices → publish → push. Publishing before the prices land builds last week's panel; pushing before publishing ships it. Nothing enforces it.

push-panel.cmd inherits cotdata's three Cygwin-rsync gotchas verbatim, since they apply unchanged: use the ssh that ships with rsync, give it a writable UserKnownHostsFile, cygdrive paths throughout.

One ordering detail that is load-bearing rather than tidy

Dated directories first, manifest.json last. The reader resolves the current week through the manifest, so a manifest naming a directory that has not landed is caught and reported as a partial sync — whereas the reverse, data present and manifest stale, silently serves last week as though it were this week. Same reason there is no latest symlink.

--delete prunes a week the publisher has already dropped from its rolling window. Unlike cotdata's push there is no consumer-owned directory on the far side to exclude, so nothing needs protecting from the mirror.

No service restart after a push. The reader caches on the manifest's mtime, so a fresh sync is picked up on the next request — unlike a cotdata store push, whose consumers hold an in-process indexer.

It is small: ~4 MB parquet plus 220 KB of blocks per week, eight weeks retained, and only the newest dated directory changes between runs.

Also fixes a claim from #67

That PR's description said the publish-location correction was "stated in all three places". The merge touched four files and bin/publish_damage.sh was not among them — the edit was made and lost before the commit, so it was two places, not three. The block is in the script now, and points at these templates rather than at an inline invocation.

Testing

560 passed / 95 skipped from a worktree, unchanged. ruff clean. bash -n clean on the modified wrapper.

Neither .cmd has been executed. I have no Windows box, and both are written by analogy to cotdata/docs/examples/windows/push-to-server.cmd. They need a real run before anyone trusts them — the likeliest snags are the %% escape in run-publish.cmd's closing comment and the .* exclusion in push-panel.cmd, which is there to keep the publisher's dot-prefixed atomic-write staging directories from propagating half-written.

🤖 Generated with Claude Code