20260915 - Draw the axes over what the node can see, not over what it saw

[Purple10101]

The Tracker page set no range on either axis, and index() rendered the template with no context at all, so the page never learned what the node was configured to see.

Autoscaling is not a neutral default here. On fairforest B this morning, 97% of the detections were a single interfering tone at 28 Hz, so the Doppler axis collapsed to a sliver around it. The plot looked full and told you nothing about the node's coverage. Two nodes, or the same node an hour apart, get drawn at different scales and cannot be compared. And when the fleet moves to ±300 Hz, the change would be invisible, because the axis was never showing ±200 in the first place.

Where the numbers come from

The node's own blah2 process.ambiguity block, via the merged config retina-gui already reads for the home, config and calibrate pages. Delay arrives in bins and becomes kilometres through capture.fs, since a bin is only a distance once the sample rate says how wide it is.

What happens when they are missing

Anything the config does not state stays None, and that axis autoscales exactly as it does today:

• no ambiguity block, so both axes autoscale
• dopplerMin without dopplerMax, so half a span is not half an axis
• delay bins with no fs, so no delay axis rather than a confident axis over a guessed cell width
• an unreadable config, so the page still renders

That last one is the point: the axes are a nicety, and losing them must not lose the plot. A guessed range misrepresents the node just as autoscaling does, only more confidently.

uirevision already preserves a viewer's own zoom across updates, so the node's span is where the axis starts rather than somewhere the view is dragged back to.

Also here: how much history is actually held

The window a viewer asks for is what they get only if the node has that much to give. It does not after a restart, and it does not once the sidecar's point ceiling starts ageing the oldest points out, which a wider Doppler span reaches sooner because it raises the detection rate.

The count line now says how much is really held when it falls short of the request, for example 4,312 detections shown · 38m of 4h held. A short record and a quiet sky are otherwise the same picture. This is computed from the data the page already has, by comparing the span it holds against the window it asked for, so it needs no change to the sidecar's stream. A precise dropped-point count would, and is not here.

Verification

ruff check clean, dead-code gate clean, 888 passed.

ruff format --check is not clean, and is not clean on main either: it wants to rewrite single to double quotes across 56 files. This repo's .pre-commit-config.yaml deliberately carries only the dead-code hook, so reformatting is out of scope for this change and would bury it.

Eight new tests in tests/test_tracker_routes.py cover the bounds being read, ±1000 Hz carried through unchanged, delay bins converted through the sample rate, and each of the four ways the bounds can be absent. The last one parses the AXES literal out of the rendered page rather than matching a substring, so it fails if the shape changes rather than only if the numbers do.

Pairs with

retina-tracker #35, which reads the same bounds on the sidecar side and holds arriving detections to them. Neither depends on the other: this one reads the node config directly, so it works against the current sidecar release.

🤖 Generated with Claude Code