TCV + TokaMark Edge-Admissibility Analytics

The event-level margin m_edge_event = C_edge_capacity − F_route_fragmentation separates all 9 reviewed TCV shots into three strictly non-overlapping states. The decisive structural result is the gap between the edge-divertor response branch (mean m = +0.660) and the three leakage branches (mean range −0.473 to −0.307). Power-balance, transport, and timing branches load the negative side exclusively; the edge-divertor response branch occupies the positive side exclusively. This is not explained by total heating power alone — three high-power ILH=1 shots still land in the negative leakage corridor, while all three positive boundary shots are defined by edge-response dominance with low transport and power-balance fragmentation scores.

The three positive boundary shots (66445, 69892, 68719) share S_edge_response ≥ 0.945 and S_transport ≤ 0.182. The four negative leakage shots have S_transport ≥ 0.960 and S_edge_response ≤ 0.287. Shot 69913 (boundary ambiguous) has the non-timing components all low (S_pb = 0.083, S_tr = 0.040, S_er = 0.019), while S_timing = 0.700 is the only elevated term — its near-zero margin reflects single-channel timing fragmentation with no compensating capacity or transport structure. Shot 68001 carries mixed ILH (1;0) and sits in the negative leakage corridor (m = −0.227), suggesting that its mixed outcome may reflect a discharge where the event-level structure was fragmentation-dominated despite eventual H-mode access.

The tag-free v0.2 margin separates 92 TCV events into three non-overlapping states with statistically significant ILH association (AUC 0.690, permutation p = 0.0022). However, 74% of events (68/92) are boundary-ambiguous — the full corpus sits near the admissibility threshold, not in the branch-rich extreme positions that characterized the 9-shot reviewed set. The UNNS margin does not improve binary ILH classification over standard plasma variables (LOO AUC 1.000 with or without m_edge_event). The correct interpretation is structural organization, not superior prediction: m_edge_event reorganizes the TCV event table into leakage-like, boundary-ambiguous, and edge-response corridors, revealing how power-balance, transport, timing, and edge-response terms oppose each other in an event-level boundary-admissibility model.

The 5-shot panel consistently ranks shot 12063 first. Its positive windows are internally coherent: C_edge_capacity > F_route_fragmentation with edge-response contribution and reduced power-balance pressure in several windows. The cross-shot comparison to 11830 confirmed reference_structurally_stronger. However, shots 11776 and 11768 showed inflated positive-boundary fractions due to missing diagnostics — a structural flaw in v0.1 that must be addressed before broader TokaMark scaling. The m_edge_conf(t) = m_edge_raw(t) × Q_diag(t) penalization scheme was defined in response.

After conservative physical relabeling and diagnostic-confidence correction, H_MODE_STABLE windows produce higher m_edge_conf medians than L_MODE windows in every shot where both labels are present. The across-label ordering is L_MODE (−1.120) < LH_TRANSITION (−0.289) < H_MODE_STABLE (+0.053). This is the ordering the UNNS-H Mode model predicts. The result passes the first gate for physical window analysis. However, it does not constitute final validation: labels are internally supported (2_SUPPORTED) but not externally anchored to expert-reviewed L-H transition times. External anchoring or a falsification stress test is required before this result can be treated as a physical validation claim.