Calibrate empirical coverage from the claim lane, not the emit loop

backend/config/constants.py

@@ -236,6 +236,39 @@ def _assoc_alt_layers_km() -> tuple[float, ...]:
 # timestamps.
 CAL_FIX_DETECTION_SKEW_S = 2.0
 
+# ── Calibration from the CLAIM lane (KNOWN_LANE_MODE != off) ─────────────────
+# Since KNOWN_LANE_MODE defaulted to "binding" (#240, 2026-08-25) the emit-loop
+# path above records nothing for a synthetic node: claiming strips the bound
+# detections from the frame before the tracker sees them, so
+# track.last_detection_adsb_hex never gets set and every synthetic node's
+# newest calibration point is dated 2026-08-25.  The claim lane is now the one
+# source under that mode (services/known_claiming.py), under a rule much
+# stricter than claiming itself.
+#
+# Residual bound for a claim that may characterize coverage, UNSCALED by fix
+# age (unlike the claim gate itself).  The claim gate is 10 µs / 25 Hz, which
+# is where it has to be to bind an echo at all; the simulator's measurement
+# noise is sigma 0.1–0.2 µs of delay and 2–4 Hz of Doppler
+# (retina_simulation.world.generate_detections_for_node), so 3 µs / 8 Hz is
+# still >5 sigma at the noisy end while shrinking the delay × Doppler area a
+# WRONG aircraft can land in by ~10x.  A claim is a bind; a calibration point
+# is a claim this node would bet its coverage polygon on.
+CAL_CLAIM_DELAY_US = 3.0
+CAL_CLAIM_DOPPLER_HZ = 8.0
+
+# Maturity bar, the counterpart of the emit path's ``n_detections >= 3``: a
+# one-frame coincidence between a wrong hex's dead-reckoned fix and a clutter
+# peak is exactly what a tight residual cannot rule out on its own, and it
+# cannot repeat frame after frame at a consistent (delay, Doppler).  Three
+# claims of the same hex by the same node, with no gap larger than
+# CAL_CLAIM_STREAK_GAP_S between consecutive ones, is a LINK rather than a
+# coincidence.  10 s: a node's frames arrive ~1 s apart and the simulator's
+# SNR-dependent miss rate reaches 40% at the detection threshold, so 2–4 s
+# gaps are routine — the same reasoning that puts KNOWN_HOLD_MAX_GAP_S at 8 s,
+# loosened because breaking a streak only costs a sample, not a track.
+CAL_CLAIM_MIN_CLAIMS = 3
+CAL_CLAIM_STREAK_GAP_S = 10.0
+
 # ── ADS-B seeding (ADSB_SEED_MODE) ────────────────────────────────────────────
 # A track view exports its ADS-B tag only if one of the newest N history
 # detections carries it.  A swapped track's newest detections go untagged

backend/core/state.py

@@ -628,6 +628,20 @@ def _adsb_for_seeding() -> dict[str, dict]:
 # what stops a node that newly acquires a silent aircraft from feeding the
 # dark pool and minting a twin key beside the lane's entry.
 known_follow_claims: int = 0
+# Empirical-coverage calibration from the CLAIM lane (see
+# services/known_claiming._calibration_from_claim and services/calibration.py's
+# fourth rule).  recorded counts the points actually written; the five rejects
+# are the five rules, charged in order, so exactly one of the six is bumped per
+# non-hold claim and they sum to the claim count.  Read them as a funnel: a
+# recorded count of zero beside a large `immature` is a fleet whose links are
+# too short-lived, and one beside a large `contested` is traffic too dense for
+# an exclusive attribution — two different problems with the same symptom.
+calibration_points_recorded: int = 0
+calibration_claims_rejected_hold: int = 0
+calibration_claims_rejected_stale_fix: int = 0
+calibration_claims_rejected_residual: int = 0
+calibration_claims_rejected_contested: int = 0
+calibration_claims_rejected_immature: int = 0
 # Dark track following (DARK_FOLLOW_MODE) — see services/dark_follow.py.
 # targets is a GAUGE (the size of the current pseudo-state list, assigned on
 # every rebuild), the other four are since-boot counters.  The funnel reads
@@ -1086,6 +1100,9 @@ def _reset_for_tests() -> None:
     global known_claims_errors, known_claims_visibility_rejects, known_claims_world_rejects
     global known_hold_claims, known_hold_expired, known_hold_dropped_disagree
     global known_follow_claims
+    global calibration_points_recorded, calibration_claims_rejected_hold
+    global calibration_claims_rejected_stale_fix, calibration_claims_rejected_residual
+    global calibration_claims_rejected_contested, calibration_claims_rejected_immature
     global dark_follow_targets, dark_follow_claims, dark_follow_inputs
     global dark_follow_published, dark_follow_dropped, dark_bottomup_shadowed
     global dark_follow_inelig_cooldown, dark_follow_inelig_no_pos
@@ -1193,6 +1210,9 @@ def _reset_for_tests() -> None:
         known_claims_world_rejects = 0
         known_hold_claims = known_hold_expired = known_hold_dropped_disagree = 0
         known_follow_claims = 0
+        calibration_points_recorded = calibration_claims_rejected_hold = 0
+        calibration_claims_rejected_stale_fix = calibration_claims_rejected_residual = 0
+        calibration_claims_rejected_contested = calibration_claims_rejected_immature = 0
         dark_follow_targets = dark_follow_claims = dark_follow_inputs = 0
         dark_follow_published = dark_follow_dropped = 0
         dark_follow_inelig_cooldown = dark_follow_inelig_no_pos = 0

backend/routes/test.py

@@ -1385,6 +1385,11 @@ def _solver_window_stats(minutes: float) -> dict:
         kh_expired = state.known_hold_expired
         kh_disagree = state.known_hold_dropped_disagree
         kf_claims = state.known_follow_claims
+        cal_recorded = state.calibration_points_recorded
+        cal_rejects = {
+            reason: getattr(state, f"calibration_claims_rejected_{reason}")
+            for reason in ("hold", "stale_fix", "residual", "contested", "immature")
+        }
         # Same one-lock snapshot for the follow lane's funnel and the
         # per-reason ineligibility tally beside it: the two are only readable
         # against each other (see the dark_follow block below), so they must
@@ -1582,6 +1587,13 @@ def _solver_window_stats(minutes: float) -> dict:
             # detections that would otherwise have started a dark twin.
             "follow_claims": kf_claims,
             "holds": sum(len(h) for h in list(state.known_track_holds.values())),
+            # Empirical-coverage calibration, which under KNOWN_LANE_MODE != off
+            # comes only from this lane (services/calibration.py's fourth rule).
+            # recorded is points written; rejected is the five rules, charged in
+            # order — exactly one per non-hold claim, so they sum with recorded
+            # to the non-hold claim count.
+            "calibration_recorded": cal_recorded,
+            "calibration_rejected": cal_rejects,
         },
         # Dark published solves against the node pool their round had for the
         # same aircraft (see the pooled/shortfalls block above).  pct is null