Stage 1 dataset evaluator ignores redundant points; re-quantify agreement precision

[jonfroehlich]

Follow-up from #9 (this part was intentionally split out).

stage_one/dataset_evaluation/evaluate.py matches one-to-one on the TP side (claims one GT, then breaks — correct), but routes redundant generated points near an already-matched GT to an "ignored" counter instead of counting them as false positives. This makes the published Stage 1 agreement precision (94.03%) slightly optimistic; recall (92.45%) is unaffected.

Fix: count those redundant points as FPs (align with rampnet/metrics.py semantics), re-run against manual_labels/, and record the delta. Expected to be small — redundant generated points within 90 px of the same ramp are rarer for Stage 1 output than for model peaks — but it should be quantified rather than assumed.

Context: README erratum and docs/eval_protocol_verification.html.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TAaPQDYhNDmYmxvrz4Lgm7

[jonfroehlich]

Closing note for whoever picks up the re-measurement: the corrected number will move for two reasons, not one.

#33 landed the code fix but deliberately did not re-run the number (it needs the generated Stage 1 dataset, which isn't in git — download_dataset.py, or run it on makelab2). In the review round it also did something this issue didn't anticipate: rather than adding a third matcher, it unified Stage 1 onto rampnet.metrics.greedy_match. So stage_one/dataset_evaluation/evaluate.py now differs from the published script in two ways:

1. Redundant points count as FPs (this issue). Lowers precision.
2. Nearest-unclaimed matching instead of first-in-list-order. These differ whenever a generated point is in range of more than one GT ramp — common: ~41% of gold-set points have a neighbour within 2× the matching radius. This can raise precision, by assigning points to ramps more sensibly and leaving fewer false positives.

The two pull in opposite directions, so the net delta from 0.9403 shouldn't be attributed to redundancy alone. Recall (0.9245) is unaffected by (1) but is affected by (2), contrary to what this issue originally said — nearest-unclaimed changes which GT points get claimed, so the TP count itself can shift.

README (erratum + the Stage 1 evaluation section) now says this too, in #42.

🤖 Generated with Claude Code

https://claude.ai/code/session_01DcegCW7sp1ZaSnR916r7QN