Velocity-consistency candidate v4 results: No-Go (4 of 558 gates)

velocity_consistency_v3 fails the frozen contract in online_pva_candidate_v4.md (committed as 5a6f48ab, code including the cherry-picked f91eb44b, before the comparison). Production defaults are unchanged; the candidate stays an explicit opt-in (--candidate velocity_consistency_v3). Machine-readable gates: online_pva_decision_v4.json (compact; the 1.6 MB full decision is in the local output directory, hash recorded there).

Control is the recorded develop c2eb06f7 default set. Candidate none built from the candidate tree is bit-identical to it in every deterministic CSV field on all 18 runs (gate 7 passes), so no control was regenerated. Replay binary SHA256 31e2859ae164c6b1a9a2bcd512778119b6420c363ed0abd71902f2f8ce045708 (also used for the contemporaneous candidate-none set), clean tree. 36 replays (six runs x normal, GNSS outage 60-70 s, IMU gap 60-64 s, control and candidate), all truth-matched, identical input hashes.

Decision

558 gates over 18 run/scenarios. Against the recorded control 21 fail: 17 are the processor P95 timing gate and 4 the GNSS-update recovery gate. The host was loaded by other jobs (load average 4-16); against the contemporaneous candidate-none run of the same binary, which is bit-identical to the control, no timing gate fails (candidate/none P95 ratio 0.92-1.61, mean 1.17; the isolated second fusion filter costs a few ms) and 4 gates fail:

Run	Gate	Control -> candidate
Tokyo 2, GNSS outage	first GNSS position update after the outage	0.0 -> 0.2 s
Tokyo 3, GNSS outage	same	0.0 -> 0.2 s
Nagoya 2, GNSS outage	same	0.0 -> 0.2 s
Nagoya 1, GNSS outage	same	0.0 -> 18.0 s

Everything else passes in all 18 run/scenarios, including what failed in v1/v2: RTK position/velocity RMSE and P95 (all-output and common-valid cohorts) are bit-identical to the control (the RTK columns below do not move), fused position, velocity and rotation RMSE/P95 are all at or below the control, coverage and first-heading/fresh-attitude times are unchanged, and the targeted rotation improvement holds on all six normal runs.

Why the recovery gate fails

The first epoch after the outage (t = 70.0 s) is an RTK FLOAT epoch. The control applies it (no gate), the candidate's position NIS gate (9 per observation) rejects it after 10 s of unaided propagation, and the next SPP epoch (0.2 s later) is accepted. Nagoya 1 is rejected for 18 s: FLOAT is rejected and the FLOAT/FIXED re-anchor (d) needs 30 consecutive FLOAT rejections. This is the cost of enabling gate (c) and of the optimism that remains in the converged FLOAT sigma (see rtk_float_covariance.md); the contract does not allow trading it against the large gain elsewhere and the candidate was not tuned after the result.

Full-run results, control -> candidate

NORMAL (control -> candidate velocity_consistency_v3)

Run	Fused pos RMSE/P95 m	Fused vel RMSE/P95 m/s	Rot RMSE/P95 deg	RTK pos RMSE/P95 m	RTK vel RMSE/P95 m/s	fused / heading avail %	first fresh/heading s
Tokyo 1	71.7 -> 14.4 / 155.2 -> 29.7	7.05 -> 0.30 / 14.62 -> 0.58	105.1 -> 1.8 / 171.1 -> 3.0	32.2 -> 32.2 / 61.7 -> 61.7	2.10 -> 2.10 / 3.29 -> 3.29	99.92/99.36 -> 99.92/99.36	2.0/15.2 -> 2.0/15.2
Tokyo 2	74.2 -> 10.3 / 179.5 -> 19.7	6.85 -> 0.28 / 14.14 -> 0.51	100.3 -> 2.0 / 170.0 -> 2.2	19.5 -> 19.5 / 29.7 -> 29.7	2.26 -> 2.26 / 3.71 -> 3.71	99.89/98.72 -> 99.89/98.72	2.0/23.4 -> 2.0/23.4
Tokyo 3	41.9 -> 14.1 / 115.0 -> 26.6	3.22 -> 0.23 / 8.22 -> 0.51	63.5 -> 1.7 / 144.4 -> 2.8	18.6 -> 18.6 / 31.9 -> 31.9	2.10 -> 2.10 / 1.76 -> 1.76	99.93/98.77 -> 99.93/98.77	2.0/37.6 -> 2.0/37.6
Nagoya 1	49.3 -> 17.7 / 118.1 -> 32.8	4.61 -> 1.02 / 9.84 -> 1.25	88.2 -> 31.1 / 168.7 -> 61.3	25.7 -> 25.7 / 29.0 -> 29.0	2.51 -> 2.51 / 4.26 -> 4.26	99.82/96.50 -> 99.82/96.50	2.0/52.0 -> 2.0/52.0
Nagoya 2	81.0 -> 22.5 / 168.0 -> 53.9	3.73 -> 0.46 / 9.96 -> 0.88	18.5 -> 2.5 / 39.8 -> 4.4	37.7 -> 37.7 / 81.8 -> 81.8	2.30 -> 2.30 / 5.09 -> 5.09	99.89/97.79 -> 99.89/97.79	2.0/41.8 -> 2.0/41.8
Nagoya 3	58.9 -> 31.9 / 128.1 -> 61.4	5.54 -> 0.35 / 11.72 -> 0.85	105.0 -> 2.9 / 170.9 -> 5.5	43.7 -> 43.7 / 101.1 -> 101.1	2.47 -> 2.47 / 5.13 -> 5.13	99.81/96.54 -> 99.81/96.54	2.0/36.0 -> 2.0/36.0

GNSS OUTAGE 60-70 s (control -> candidate velocity_consistency_v3)

Run	Fused pos RMSE/P95 m	Fused vel RMSE/P95 m/s	Rot RMSE/P95 deg	RTK pos RMSE/P95 m	RTK vel RMSE/P95 m/s	fused / heading avail %	first fresh/heading s
Tokyo 1	55.1 -> 14.1 / 130.9 -> 27.9	6.11 -> 0.28 / 11.78 -> 0.56	104.9 -> 1.7 / 171.2 -> 2.8	29.7 -> 29.7 / 58.9 -> 58.9	2.64 -> 2.64 / 3.89 -> 3.89	99.92/99.36 -> 99.92/99.36	2.0/15.2 -> 2.0/15.2
Tokyo 2	72.5 -> 8.5 / 187.3 -> 18.9	4.82 -> 0.27 / 11.98 -> 0.63	28.5 -> 2.6 / 68.1 -> 2.5	19.4 -> 19.4 / 26.4 -> 26.4	2.42 -> 2.42 / 4.99 -> 4.99	99.89/98.72 -> 99.89/98.72	2.0/23.4 -> 2.0/23.4
Tokyo 3	27.2 -> 14.4 / 31.9 -> 26.7	1.95 -> 0.22 / 1.40 -> 0.48	12.7 -> 1.6 / 40.6 -> 2.9	16.8 -> 16.8 / 30.9 -> 30.9	1.07 -> 1.07 / 1.55 -> 1.55	99.93/98.77 -> 99.93/98.77	2.0/37.6 -> 2.0/37.6
Nagoya 1	53.2 -> 18.4 / 131.9 -> 36.5	4.87 -> 1.13 / 10.91 -> 1.30	88.7 -> 42.6 / 168.8 -> 146.0	26.0 -> 26.0 / 29.3 -> 29.3	2.76 -> 2.76 / 4.39 -> 4.39	99.82/96.50 -> 99.82/96.50	2.0/52.0 -> 2.0/52.0
Nagoya 2	55.2 -> 13.2 / 133.2 -> 30.9	4.30 -> 0.38 / 9.76 -> 0.70	90.8 -> 2.0 / 167.5 -> 2.8	37.2 -> 37.2 / 85.7 -> 85.7	4.24 -> 4.24 / 13.91 -> 13.91	99.89/97.79 -> 99.89/97.79	2.0/41.8 -> 2.0/41.8
Nagoya 3	75.2 -> 31.7 / 171.7 -> 61.4	5.75 -> 0.36 / 10.27 -> 0.85	103.2 -> 2.8 / 171.3 -> 5.4	45.8 -> 45.8 / 103.7 -> 103.7	2.47 -> 2.47 / 6.22 -> 6.22	99.81/96.54 -> 99.81/96.54	2.0/36.0 -> 2.0/36.0

Recovery s (GNSS update / fresh attitude / heading), control -> candidate: - Tokyo 1: 0.0 / 0.0 / 0.0 -> 0.0 / 0.0 / 0.0 - Tokyo 2: 0.0 / 0.0 / 0.0 -> 0.2 / 0.0 / 0.0 - Tokyo 3: 0.0 / 0.0 / 0.0 -> 0.2 / 0.0 / 0.0 - Nagoya 1: 0.0 / 0.0 / 0.0 -> 18.0 / 0.0 / 0.0 - Nagoya 2: 0.0 / 0.0 / 0.0 -> 0.2 / 0.0 / 0.0 - Nagoya 3: 0.0 / 0.0 / 0.0 -> 0.0 / 0.0 / 0.0

IMU GAP 60-64 s (control -> candidate velocity_consistency_v3)

Run	Fused pos RMSE/P95 m	Fused vel RMSE/P95 m/s	Rot RMSE/P95 deg	RTK pos RMSE/P95 m	RTK vel RMSE/P95 m/s	fused / heading avail %	first fresh/heading s
Tokyo 1	63.4 -> 15.0 / 148.0 -> 32.2	5.83 -> 0.29 / 11.39 -> 0.61	104.9 -> 1.8 / 171.3 -> 3.1	31.2 -> 31.2 / 63.3 -> 63.3	2.36 -> 2.36 / 3.72 -> 3.72	99.67/98.39 -> 99.67/98.39	2.0/15.2 -> 2.0/15.2
Tokyo 2	52.2 -> 11.8 / 111.5 -> 20.1	6.07 -> 0.40 / 12.06 -> 0.93	102.7 -> 18.4 / 170.6 -> 55.6	14.0 -> 14.0 / 23.6 -> 23.6	1.66 -> 1.66 / 2.21 -> 2.21	99.57/97.89 -> 99.57/97.89	2.0/23.4 -> 2.0/23.4
Tokyo 3	59.3 -> 14.1 / 90.2 -> 26.7	6.60 -> 0.22 / 12.34 -> 0.48	104.6 -> 2.0 / 171.2 -> 3.3	17.3 -> 17.3 / 31.0 -> 31.0	1.53 -> 1.53 / 1.74 -> 1.74	99.75/98.57 -> 99.75/98.57	2.0/37.6 -> 2.0/37.6
Nagoya 1	55.8 -> 17.6 / 130.6 -> 33.0	4.91 -> 0.96 / 11.22 -> 1.07	83.0 -> 14.1 / 167.4 -> 16.5	27.6 -> 27.6 / 30.8 -> 30.8	2.57 -> 2.57 / 4.34 -> 4.34	99.43/96.09 -> 99.43/96.09	2.0/52.0 -> 2.0/52.0
Nagoya 2	64.2 -> 15.1 / 146.2 -> 32.7	4.41 -> 0.34 / 9.38 -> 0.70	104.1 -> 2.3 / 170.7 -> 4.6	38.3 -> 38.3 / 85.0 -> 85.0	4.34 -> 4.34 / 14.59 -> 14.59	99.59/97.46 -> 99.59/97.46	2.0/41.8 -> 2.0/41.8
Nagoya 3	88.6 -> 32.8 / 192.6 -> 61.4	7.26 -> 0.41 / 14.31 -> 0.87	106.7 -> 2.8 / 170.3 -> 5.4	43.8 -> 43.8 / 100.7 -> 100.7	2.04 -> 2.04 / 3.82 -> 3.82	99.25/95.94 -> 99.25/95.94	2.0/36.0 -> 2.0/36.0

Recovery s (GNSS update / fresh attitude / heading), control -> candidate: - Tokyo 1: 2.0 / 2.0 / 19.4 -> 2.0 / 2.0 / 19.4 - Tokyo 2: 2.0 / 2.0 / 11.4 -> 2.0 / 2.0 / 11.4 - Tokyo 3: 2.0 / 2.0 / 2.4 -> 2.0 / 2.0 / 2.4 - Nagoya 1: 2.0 / 2.0 / 2.4 -> 2.0 / 2.0 / 2.4 - Nagoya 2: 2.0 / 2.0 / 2.4 -> 2.0 / 2.0 / 2.4 - Nagoya 3: 2.0 / 2.0 / 2.4 -> 2.0 / 2.0 / 2.4

Availability is unchanged to the printed precision. Nagoya 1 rotation is still 31-43 deg (reverse start, not addressed by any candidate) but is better than the control's 88 deg; Tokyo 2 IMU-gap rotation 102.7 -> 18.4 deg is the only remaining non-converged gap case besides Nagoya 1.

Ablation used to choose the components (development, normal runs, not gated)

Fused position RMSE (m) / rotation RMSE (deg). iso = isolated RTK prior. Covariance mode 3 everywhere except none.

Variant	Tokyo 1	Tokyo 2	Tokyo 3	Nagoya 1	Nagoya 2	Nagoya 3
control	71.7 / 105	74.2 / 100	41.9 / 63	49.3 / 88	81.0 / 18.5	58.9 / 105
covariance only	36.2 / 105	19.6 / 7.6	16.3 / 8.1	23.3 / 106	23.8 / 4.2	31.9 / 5.1
+ c	12.5 / 1.8	10.0 / 3.0	14.3 / 1.9	54.0 / 32	46.1 / 3.2	32.3 / 2.9
+ a,c	12.2 / 1.9	10.0 / 3.0	13.8 / 1.9	54.0 / 37	17.3 / 3.0	32.3 / 3.5
+ a,b,c	25.5 / 3.2	7.8 / 1.7	13.3 / 1.7	118 / 45	16.3 / 4.0	28.9 / 2.8
+ a,b,c,d	20.3 / 2.4	6.4 / 1.7	5.2 / 1.6	29.7 / 40	14.5 / 4.0	26.5 / 2.5
iso + c	14.7 / 1.8	11.6 / 2.0	14.4 / 1.7	54.6 / 31	38.2 / 2.2	31.9 / 2.9
iso + c,d (candidate)	14.4 / 1.8	10.3 / 2.0	14.1 / 1.7	17.7 / 31	22.6 / 2.5	31.9 / 2.9

Reading: the honest covariance alone removes the attitude collapse on three runs (rotation 63-105 -> 5-8 deg on Tokyo 2, Tokyo 3, Nagoya 3; Nagoya 2 18.5 -> 4.2) and roughly halves position error, but not Tokyo 1 and Nagoya 1. Without isolation the RTK columns move (without b: Tokyo 2 RTK position 19.5 -> 17.7, Tokyo 3 18.6 -> 21.0 with a, Nagoya 2 +15 % with covariance alone, RTK velocity 2.10 -> 2.8 on Tokyo 1), because the fused attitude bootstraps the RTK prior; (b) moves them further both ways. With isolation they are identical by construction. Component (a) is neutral to harmful once the covariance is honest; (d) matters for Nagoya 1 position only.

Limits and unverified

• Isolation (f) forgoes any RTK benefit from the better fused state; whether the RTK filter should consume the corrected state was not evaluated and would need its own RTK-gate decision.
• The remaining optimism of the converged FLOAT sigma (RMS z about 4-6) is not fixed; the post-outage rejection above follows from it.
• Nagoya 1 reverse start and Tokyo 2 IMU-gap rotation remain.
• Only the six development runs; no holdout. Timing is host-load dependent; the contemporaneous-none comparison is the meaningful one and it is the same binary, so it measures only the candidate's own cost.
• The ablation grid above used a scratch instrumented copy of the replay with environment switches (not committed). Its iso + c,d normal-run output is bit-identical, in every deterministic CSV field, to the committed velocity_consistency_v3 normal-run output on all six runs.

Validation

New unit tests: RTKCovarianceConsistencyTest (4 cases: consistent FLOAT is not inflated; inflated FLOAT reaches the 3-dof statistic exactly and reports a sigma comparable to the disagreement; scale grows with disagreement and shrinks with SPP uncertainty; invalid input reports no evidence), the OFF defaults of the new RTK and online options, and the comparator test for the new candidate name. The v2 FloatGateRecovery* fusion tests come with the cherry-picked recovery. Release build, Python bindings off: run_tests 1,583 cases, 1,516 pass, 67 skipped (missing fixtures), 0 failed; gnss_online_tests 13/13; test_pva.py and test_pva_comparison.py pass. The RTK-level behaviour of the modes is covered by the 36-replay and batch comparisons above, not by a unit test (the bundled kinematic fixtures are absent in this checkout). Broader ctest lanes were not run.