#1520
vsolve on: two passes skipped, snap after dips and carriageway ends (same or better)
Stack (merge in this order): #1418 → #1426 → #1465 → #1470 → #1517 → #1520. Each PR's diff against
devholds the ones before it; this one's own changes: 7aeae0df, da12617a and the merge fcff0c8a.
The RFC's step 4, second half: remove the passes the one solve replaces. Each pass was skipped alone in on mode (BNE_VSOLVE_SKIP=<pass>) on the 8 baseline places.
What
- In on mode, two passes are no longer run: the junction snap after the dips (
snap_dips) and the carriageway ends' pin (cw_ends). List:vsolve.SKIP.BNE_VSOLVE_SKIP=noneruns every pass again. - Where the solve finds no heights, on mode now falls back like solo mode does: the plans go back to the profiles before the passes, and every pass runs again. Test:
SoloTest.test_on_mode_s_skipped_passes_come_back_without_a_solve. - Plans key:
.vs1.k1(vsolve.SKIP_TAG). Another list gets its own tag.
Numbers
| pass skipped alone | what moved (on → skipped) |
|---|---|
neighbours, coarse, guard, struct_ends | nothing, because they are off by default |
cw_ends | nothing |
snap_dips | only better: Tokyo made steep 104 → 103, corridor 819 → 813; Rio corridor 184 → 181 |
ramp, snap_weld, weld, under | a little worse (Tokyo corridor +2 to +7, Athens ground p95 +0.01) |
snap, corridors | clearly worse (Tokyo corridor 889 / 905, Athens ground p95 1.65) |
With both skipped:
- grade_metrics: the same as
snap_dipsalone, so the same or better on every metric; - test_drivability: 1277.0. That is the same as #1517: every site unchanged by the skip.
CPU: these runs can't show a gain this small. The base moved from 490 s to 511 s between two runs. The solo result already said that removing the passes is not where the CPU goes.
Details: docs/research/2026-11-21-vsolve-skip.md.