B1: /api/cma/dashboard/predict hand-rolled threshold parsing called float() on the whole
expression -> ValueError on multi-clause thresholds like '<=36|>=124' -> 500. Now routes
through app.utils.threshold_parser.eval_threshold (single source of truth; red then yellow,
invalid/blank expr -> False, no crash).
B2: ReportCenter.vue restatement tab template called restateDiffClass while the script
defines restatementDiffClass (name mismatch since ec6af75) -> TypeError on render.
Verified: predict 200 on new code vs 500 on live code; /reports A/B browser run shows 0 vs 3
occurrences of the TypeError; pytest 1113 passed / 1 xfailed; vite build OK.
P-6 (M2 gate 2, FIN-CE-REVIEW-20260915-01): kpis.REVERSE_INDICATORS is now the only authoritative direction definition; dashboard.py imports it instead of carrying a literal copy; bot_kpis.py derives REVERSE_BOT_INDICATORS from it (dead import removed).
F_FACTORY_REBATE_RATE (upstream factory rebate rate, higher=better) was wrongly treated as reverse: 278.7% vs target 80 scored 1/danger, now 5/success and workbench three-color red -> green. No DB write, no scoring-engine logic change, no data_level/frontend touch.
Single source: app/utils/voucher_guard.py (NOT EXISTS on voucher_void_marks,
key = entity_id+voucher_no+voucher_date and revoked_at IS NULL; same caliber as
the DB view cma.v_voucher_details_active). No hardcoded voucher numbers.
Consumers now guarded: cash.py import linkage (2 fns), reports.py subject/BS
amount (2 sites, also entity_id isolation added), dashboard.py _bep_sum_field,
alert_attribution.py build_subject_attribution. kpi_value_collector.py now
delegates to the shared guard (removes the previous duplicate implementation).
Tests: new tests/test_void_marker_consumers_20260914.py (6 cases).
Hermes CI Fix