nx_wfcost_outcomes_t144.nx

nx_wfcost.nx -- WORKFLOW-EPISODE TOKEN COST (Lane I, common-task hardening loop, 2026-08-20). Operator: "grow our capabilities into repeatable high value little to no token capabilities". Ranking hardening candidates needs the REAL cost of what we already do repeatedly. The estate had the three halves and never joined them: nx_actlog mines recurring tool:verb support, nx_eff_lib measures transcript tokens PER DAY, and nx_tokroi's uses_per_day is a DECLARED assumption. This organ joins them: per-INVOCATION-PAIR measured token cost over the FULL transcript population. THE CORRECTION THIS ORGAN CARRIES (measured 2026-08-20, and it is why the join could not be trusted before): one assistant MESSAGE is written to the transcript as SEVERAL JSONL LINES -- one per content block (thinking, tool_use, text) -- and EVERY one of those lines repeats the SAME usage object verbatim. nx_eff_lib::es_tr_tokens dedups the two usage copies WITHIN a line (message.usage and the iterations[] echo -- that half is correct and gated by its T20) but counts once PER LINE, so a message with N content blocks is counted N times. Measured over the 12 largest transcripts: 43,372 assistant lines carry only 22,234 distinct message ids => the per-line method overstates by ~1.95x. This organ dedups by message.id through a real hash set and PRINTS BOTH TOTALS side by side so the correction is proven by the run, never asserted (two counters of one population = free audit). Dedup is a SET, not a previous-id compare: 12 of 22,234 ids reappear NON-consecutively, so the cheap compare would silently overcount those. gauge [trdir] -> corrected(msgid-deduped) vs perline totals + overcount permil + envelope mine [trdir] [minsup] -> recurring invocation PAIRS >= minsup with exact total measured tokens selftest -> gv_verdict gate, named neg-controls, runtime fixtures RANKING QUANTITY IS THE EXACT TOTAL, NOT A MEDIAN: total = the sum of every episode's measured cost, carried exactly. A median would need a second pass plus per-episode retention; mean and max are printed and LABELLED as such, and no median is claimed. license_tier: ORIGINAL No hw writes (Rule 26).