@nexus-cortex/cli 4.124.16 → 4.124.18

package/.env.defaults

@@ -1172,6 +1172,40 @@ CORTEX_ENDTURN_GATE=true
 # (no lever) R187c (4.124.15): the requirement-ledger extractor gets the planner's timeout (180 s, CORTEX_JUDGE_REQ_LEDGER_TIMEOUT_MS
 # override) instead of the judge's 90 s (a 98-s pro call lost the whole ledger in the 9:04 AM L7 smoke), and an empty lift-time result is
 # re-authored ONCE at the first finish.
+# (no lever) R191 HB-LEDGER-EXTRACTOR-HYGIENE (4.124.17, 2026-09-23; L7 interim read, 5 sessions): (a) the sglang ledger carried "The DOM
+# must not change" — copied from the persona's own example list — and the writer called it bizarre; (b) 4 of 5 ledgers spent 1–2 of their 8
+# lines on conduct text ("complete within 28800 seconds", "do not cheat / no online hints"); (c) photonic's extractor guessed a checker's CLI
+# (`check_routing.py <spec> <result>` vs the documented `--layout 1`), two lines FAILED at every finish on a PASSING solution, and the writer
+# spent turns on the "harness quirk". Fix: persona lists kinds abstractly, marks the example as from an unrelated task, skips conduct lines,
+# and invokes a named checker exactly as documented; the parser drops conduct lines (isBoilerplateRequirement); a check run whose failure is
+# the CHECK's own defect (usage/argparse error, missing script or tool, SyntaxError/NameError) is BROKEN → the line becomes UNVERIFIABLE with
+# its check dropped (never re-run, never a failure, Jev-decidable); the endturn_resolver event now banks reqBroken + reqLinesDetail (per-line
+# status/runs/last result) so the false-failure rate is readable from the decisions file. Interim L7 (5/21 rows): ledger authored + delivered
+# 5/5, 8 lines each, 1–6 with checks; reqOpen was 0 at every finish (the harness exercises every check itself, so the OPEN hold never fires;
+# the ledger's live effects are FAILED-as-evidence and Jev closes); every passing session had ≥1 FAILED line at its accepted finish.
+# CORTEX_JUDGE_REQ_LEDGER_FAIL_VETO / _MAX / _MIN_REMAINING — R192 HB-LEDGER-FAILED-VETO (4.124.18, 2026-09-23; operator: "build the
+# ledger-failed veto with jev arbitration in shadow"). L7 (15 rows): every failing consistency session ended accept-with-gap with 45–92% of
+# the budget left while a ledger line whose check the harness had JUST run was FAILED on the real defect (photonic R5 separation, telecom R8)
+# — the evidence existed and the acceptance policy accepted over it (one shared evidence veto per session, then accept-with-gap). With
+# FAIL_VETO=on a would-be accept-with-gap with a line FAILED at THIS finish is returned to the writer with the failing check + output, at most
+# _MAX (2) times per session, never below _MIN_REMAINING (0.25) of the budget. Precedence: judge veto > ledger-failed veto > R187 hold >
+# derivation hold. Narrower than R173 GAP_HOLD (g1 11 vs 9, g2 8 vs 9 — noise): it triggers only on an objective failing check for a STATED
+# requirement, never on judge prose. Unverifiable lines (sglang's class) are out of reach by design. Default off.
+# [LEDGER 2026-09-23 @4.124.18] BUILT + unit-tested (R192 block in requirementLedger.test.ts, 4 tests). FIELD: pending — the L7 shape
+# (7 consistency tasks ×3, ~$14) with FAIL_VETO=on + FAIL_JEV=shadow, paired against tb2-t4l7-ctl; bar = the three FAILED-at-accept
+# sessions convert without regression.
+#CORTEX_JUDGE_REQ_LEDGER_FAIL_VETO=
+#CORTEX_JUDGE_REQ_LEDGER_FAIL_VETO_MAX=
+#CORTEX_JUDGE_REQ_LEDGER_FAIL_VETO_MIN_REMAINING=
+# CORTEX_JUDGE_REQ_LEDGER_FAIL_JEV / _MIN — R192 Jev arbitration of a FAILED line BEFORE the veto: one noul per line ("is this failure a
+# defect of the CHECK itself — wrong args/path/logic — rather than the work?") on the requirement, the check, its output and the writer's
+# attestation (the typed-gate-on-concrete-evidence role Jev scores 0.98 on; never a doneness call). off | shadow (default: ask + bank
+# `reqFailJev {asked, defects, probs}` in the endturn_resolver event, act on nothing) | on (p ≥ _MIN 0.7 → the line becomes UNVERIFIABLE,
+# check dropped, and leaves the veto set). Runs only when FAIL_VETO is on and a line failed at this finish (~0.4 s, pennies).
+# [LEDGER 2026-09-23 @4.124.18] BUILT + unit-tested. FIELD: shadow in the R192 cell; replay Jev's defect calls against the graders
+# (a line called "defect" in a PASSING session = correct; in a FAILING session on the real requirement = wrong) before ever setting `on`.
+#CORTEX_JUDGE_REQ_LEDGER_FAIL_JEV=
+#CORTEX_JUDGE_REQ_LEDGER_FAIL_JEV_MIN=
 # CORTEX_JUDGE_INDEPENDENT_DERIVATION / _TOL — R176 HB-INDEPENDENT-DERIVATION (4.122.0, 2026-09-21; design docs/R176_INDEPENDENT_DERIVATION_DESIGN.md).
 # The largest never-passed TB4.0 class (~14 tasks: foodstuff, roy-polymorph, glycan, gsea, hof, vigenere, sound-change, cargo, ctr,
 # shadow-relay, satb, takens…) is a computed value / exact artifact graded against a golden answer with NO ground truth in the container:

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@nexus-cortex/cli",
-  "version": "4.124.16",
+  "version": "4.124.18",
   "description": "Nexus Cortex CLI - Terminal interface for multi-provider LLM orchestration",
   "main": "dist/index.js",
   "type": "module",
@@ -19,7 +19,7 @@
     "prepack": "node ../../scripts/copy-pkg-cortex-scaffold.mjs"
   },
   "dependencies": {
-    "@nexus-cortex/core": "4.124.16",
+    "@nexus-cortex/core": "4.124.18",
     "chalk": "^4.1.2",
     "cli-spinners": "^2.9.0",
     "commander": "^11.0.0",