Test overlay: run the measured dark-solver configuration on the test droplet

backend/.env.example

@@ -121,17 +121,69 @@ MENDER_PAT=
 # SOLVER_RESOLVE_INTERVAL_S is how long an aircraft is not re-solved for.
 # Association is rate-limited per node, so every node that sees an aircraft
 # emits its own candidate for it and most arrivals are duplicates; this is the
-# window that suppresses them. Do not exceed the 60 s multinode_tracks expiry —
+# window that suppresses them. Do not exceed the multinode_tracks expiry
+# (MN_DARK_EXPIRY_S below for dark tracks, 60 s for ADS-B-assisted ones) —
 # tracks would lapse between solves. 0 turns the suppression off. See
 # services/tasks/solver.py and docs/runbook.md's solver_latency_high entry.
 #
 # SOLVER_MAX_DISPLACEMENT_KM_DARK is how far a DARK (no-ADS-B) solve may land
 # from its association guess before it is rejected; the default is 6 km = 2 x
 # the 3 km association grid step, since a dark guess is a quantised grid point
 # rather than an ADS-B fix. ADS-B-anchored solves keep the fixed 2 km cap.
+#
+# MN_DR_CAP_S is how far a multinode entry is dead-reckoned past its last
+# solve before it holds position, and MN_DARK_EXPIRY_S is when a DARK
+# (mn-dark-*) entry is dropped from the feed entirely. Both are position-error
+# budgets: measured against ground truth, dark entries run 1.5 km median error
+# at 8-15 s of solve age but 2.0 km at 15-30 s and 4.0 km at 30-60 s. ADS-B-
+# assisted (mn-adsb-*) entries keep the 60 s expiry — a transponder hex anchors
+# them. Raise both to trade position accuracy for track continuity.
+#
+# SOLVER_ALT_MODE is how an n>=3 solve gets its altitude. sweep (default)
+# solves once per fixed altitude layer and keeps the lowest rms_delay — six
+# process-pool round trips per candidate, and an altitude quantised to a
+# ladder whose 2 km spacing puts up to 1 km of error straight into the
+# residual the reject gate reads. free makes one pool call to the geolocator's
+# multi-start helper, which solves altitude as a sixth unknown. Both modes
+# stamp altitude_mode on the solve-history record, so /api/test/mlat-history
+# can compare them across a deploy of each. See docs/solverflow.md.
+#
+# SOLVER_FREE_ALT_STARTS is how many start altitudes free mode gives that
+# helper, clamped to the number of layers. 1 (the default) starts at the layer
+# nearest the association guess, or at the guess altitude itself when it comes
+# from ADS-B; more is a window around it, at one LM run each. Measured on test
+# over 1019 free solves, three starts changed rms_delay by more than 0.1 us in
+# 13 of them, so the extra runs are off by default and worth turning on only
+# where the geometry sends a single start to the wrong side of an ellipse.
 # SOLVER_WORKERS=2
 # SOLVER_RESOLVE_INTERVAL_S=12
 # SOLVER_MAX_DISPLACEMENT_KM_DARK=6.0
+# MN_DR_CAP_S=15.0
+# MN_DARK_EXPIRY_S=30.0
+# SOLVER_ALT_MODE=sweep
+# SOLVER_FREE_ALT_STARTS=1
+
+# Dark track following (services/dark_follow.py, docs/solverflow.md 3d). The
+# known lane's top-down claim applied to aircraft with no transponder: an
+# established mn-dark-* track's Kalman state predicts delay/Doppler at each
+# node, matching detections are claimed against it, and the resulting solve
+# carries the followed key as its anchor so it lands back on the SAME track.
+#
+# DARK_FOLLOW_MODE: off | shadow | binding. shadow (the default) claims,
+# solves and records but leaves the detections in the dark pool and publishes
+# nothing; binding strips them and publishes through the normal dark gates.
+#
+# The rest are the eligibility and guard thresholds. A key is followable while
+# its last solve is under MAX_AGE_S old with at least MIN_SOLVES solves and a
+# filter velocity sigma under MAX_VEL_SIGMA_MS; it is dropped for COOLDOWN_S
+# after two consecutive rejected follow-solves, which is what stops a ghost
+# locking itself onto the map. INTERVAL_S rate-limits follow-solves per key.
+# DARK_FOLLOW_MODE=shadow
+# DARK_FOLLOW_MAX_AGE_S=20
+# DARK_FOLLOW_MIN_SOLVES=3
+# DARK_FOLLOW_MAX_VEL_SIGMA_MS=60
+# DARK_FOLLOW_INTERVAL_S=2.0
+# DARK_FOLLOW_COOLDOWN_S=30
 
 # Detection mirror. Production only. Every accepted v1 detection frame is
 # forwarded to another environment's /api/radar/detections/bulk, batched once a

backend/config/constants.py

@@ -84,11 +84,25 @@ def as_num(v) -> float:
 N2_CONFIRM_MIN_EPOCHS = 4  # Floor on samples; span is the real gate
 N2_TRACK_HISTORY_MAX = 20  # Per-node track samples fed to the fit
 
+# How old a track's newest REAL detection may be before the track stops being
+# offered to association (see services/frame_processor.confirmed_track_views).
+# A COASTING track is kept alive for N_DELETE=10 frames after its last
+# association, and at the fleet's 0.74-1 Hz per-node cadence that is up to ~13 s
+# of dead reckoning.  confirmed_track_views hands association the track's last
+# real sample, and association hands the solver that sample as if it were
+# current — so an aircraft that has flown out of a node's beam keeps
+# contributing a seconds-old delay to n>=3 solves.  Measured on the test
+# droplet: 230 out of-cone nodes survived into published dark solves in 20 min,
+# median 4 deg outside the beam edge (p90 22 deg) and 1.8 km beyond max range,
+# and they are the nodes the rms trim then throws away.  0 disables the filter.
+TRACK_MAX_STALE_S = float(os.getenv("TRACK_MAX_STALE_S", "3.0"))
+
 # A 2-node track needs this many solves before it renders a plane; 1
 # disables the gate.  One-shot n=2 solves were the dominant ghost source.
 MN_N2_MIN_SOLVES = int(os.getenv("MN_N2_MIN_SOLVES", "2"))
 # One-shot display lifetime for n>=3 solves, seconds.  A track confirmed by
-# a second solve gets the normal 60 s entry expiry / 30 s DR cap.
+# a second solve gets the normal MN_DARK_EXPIRY_S/60 s entry expiry and the
+# MN_DR_CAP_S dead-reckoning cap below.
 #
 # The window has to outlive the wait for the confirmation it is waiting for,
 # or it is not a preview window — it is a guaranteed disappearance.  A second
@@ -97,11 +111,36 @@ def as_num(v) -> float:
 # cadence is a 9 s median and a 25 s p90.  At 5 s the great majority of
 # genuine n>=3 one-shots blinked out before the round that would have
 # confirmed them ever ran, which reads on the map as flicker, not as caution.
-# 15 s covers the median and most of the p90 while still being well short of
-# the 60 s entry expiry, so an unconfirmed one-shot is still withdrawn long
-# before a confirmed track would be.
+# 15 s covers the median and most of the p90 while still being short of the
+# entry expiry (MN_DARK_EXPIRY_S 30 s dark, 60 s assisted), so an unconfirmed
+# one-shot is still withdrawn before a confirmed track would be.
 MN_ONESHOT_TTL_S = float(os.getenv("MN_ONESHOT_TTL_S", "15.0"))
 
+# How far a multinode entry may be dead-reckoned past its last solve, seconds.
+# Beyond this it holds its last dead-reckoned point until the entry expires.
+#
+# The cap is a position-error budget, not a cadence allowance.  Measured on
+# the test droplet over 20 minutes (dark multinode feed entries vs ground
+# truth): median error 1.05 km under 3 s of solve age, 1.21 km at 3–8 s,
+# 1.50 km at 8–15 s, then 2.02 km at 15–30 s (7% of entries more than 5 km
+# off).  The knee is at 15 s, which is where the KF's learned velocity error
+# starts to dominate the solve error it is extrapolating.  The old 30 s cap
+# was set when a dark aircraft was re-solved every 12 s and the extra window
+# bought coverage; dark solves now land every 1–3 s, so a 15 s gap is a lost
+# track rather than a cadence gap, and extrapolating it only invents motion.
+MN_DR_CAP_S = float(os.getenv("MN_DR_CAP_S", "15.0"))
+
+# Entry expiry for DARK multinode tracks (mn-dark-*), seconds.  ADS-B-assisted
+# entries (mn-adsb-*) keep the 60 s expiry: they are anchored to a transponder
+# fix, so a gap there is the ADS-B feed breathing rather than a lost target.
+#
+# Same 20-minute capture: the 30–60 s age band was 12% of all displayed dark
+# entries, with a 3.99 km median error and 32% more than 5 km off — an icon
+# that reads as a live target while sitting kilometres from any aircraft.  At
+# the current 1–3 s dark solve cadence an entry that has not re-solved in 30 s
+# is a lost track, and withdrawing it is more honest than holding it.
+MN_DARK_EXPIRY_S = float(os.getenv("MN_DARK_EXPIRY_S", "30.0"))
+
 # Quality gate for adopting the constant-velocity fit's velocity into a
 # published solve, in place of the single-epoch Doppler solution (see
 # solver.py's _resolve_cv_fit / velocity adoption in _process_solver_item).