feat: audit.py computes I1, I5 (narrow), B1, B4, B5 and F5 from a real workbook

.githooks/pre-commit

@@ -106,4 +106,16 @@ if git diff --cached --name-only | grep -qE '^(test/.*\.py|skills/trz-expert/scr
     fi
     rm -f /tmp/trz-kchecker.$$
     echo "pre-commit: K5/K6 checker matches the manifest"
+
+    echo "pre-commit: B/F/I checker against generate_wide.py's manifest, 25 seeds"
+    if ! "$PY" test/audit_test.py --seeds 25 > /tmp/trz-audit.$$ 2>&1; then
+        cat /tmp/trz-audit.$$
+        rm -f /tmp/trz-audit.$$
+        echo
+        echo "Commit stopped: audit.py disagrees with generate_wide.py's manifest "
+        echo "or expected_findings.md. To skip: git commit --no-verify"
+        exit 1
+    fi
+    rm -f /tmp/trz-audit.$$
+    echo "pre-commit: B/F/I checker matches the manifest"
 fi

.github/workflows/tests.yml

@@ -93,6 +93,13 @@ jobs:
         # what would have caught it.
         run: python test/k_checker_test.py --seeds 300
 
+      - name: B/F/I computed from a workbook, against generate_wide.py's manifest
+        # scripts/audit.py's own hand-built fixtures for I5/B5, the suite-1 static
+        # fixture checked against expected_findings.md (including the part-time row
+        # that must stay silent - the exact row that caught a false positive in this
+        # script before anyone else did), plus 300 seeds for I1/B4/F5.
+        run: python test/audit_test.py --seeds 300
+
       - name: Suite 5 - the chain below the payroll
         # ведомост -> обр. 1 -> обр. 6 -> внесено, one link broken at a time. The
         # checks that reconcile the payroll against the documents downstream of it had
@@ -315,3 +322,5 @@ jobs:
         run: python test/run_tests.py --seeds 3000
       - name: K5/K6 checker at the same depth
         run: python test/k_checker_test.py --seeds 3000
+      - name: B/F/I checker at the same depth
+        run: python test/audit_test.py --seeds 3000

CLAUDE.md

@@ -34,6 +34,7 @@ python test/checks_test.py    # suite 1: static payroll against the key in expec
 python test/eval_skill.py --selftest      # free: checks the refusal grading itself
 python test/preflight_test.py # scripts/preflight.py: clean is silent, each shape defect found once
 python test/k_checker_test.py # scripts/k_checker.py: K5/K6 vs. generate_wide.py's manifest, 60 seeds by default
+python test/audit_test.py     # scripts/audit.py: I1/B4/F5 vs. the manifest, B1/I5/B5 vs. hand-built and suite-1 fixtures
 python test/komplekt_test.py  # suite 5: ведомост -> обр. 1 -> обр. 6 -> внесено -> счетоводство, one link at a time
 python test/lifecycle_test.py # suite 6: five months of the same people, one timeline break at a time
 python test/run_tests.py      # all five, 50 seeds
@@ -128,6 +129,24 @@ A false positive fails exactly like a miss.
   `preflight.py`'s own `SKILL_DIR`-relative path to `stavki.md`, and the pre-commit
   hook's trigger regex - a change under `scripts/` that stops matching that regex would
   go untested locally again, the exact failure mode the move was meant to close.
+- **`scripts/audit.py` (2.18.0) covers I1, I5 (narrow), B1, B4, B5 and F5 - not the
+  rest of B/F/I/K.** Each is mechanical (a row's own numbers, or a rate read fresh
+  from `references/stavki/` via `scripts/rates.py`, never typed into either file) and
+  safe for a generic tool - B2/B3/B6 need a company-specific number `mapping.yaml`
+  doesn't carry, F1/F2/F3/F4/F6/F7/F9's composition method needs a judgment call this
+  script isn't positioned to make safely yet, F8/F10 need information one workbook
+  doesn't hold, and K1/K3/K4/K7/K8 are `k_checker.py`'s same closed-vocabulary risk.
+  Building it against `generate_wide.py` at scale (not just a hand-built fixture)
+  found a real, pre-existing `preflight.py` vocabulary bug before it ever shipped:
+  "НЕТО преди удръжки"/"НЕТО за изплащане" and "Вноски работодател ДОО+ТЗПБ"/"Вноски
+  работодател общо" both collapsed into one concept each, raising the blocking
+  `DUPLICATE_CONCEPT` signal on every realistic fixture and stopping every check cold -
+  split now, the same way "Клас %"/"Клас сума" already were. B1/B5 also went through
+  two rounds of false positives from partial attendance (part-time hours, then a
+  partial month from leave/sick days) before landing on "the highest count declared
+  on the sheet" as the least-wrong stand-in for the full-time/full-month norm this
+  script has no public-holiday calendar to compute directly - the suite-1 fixture's
+  own part-time row (Стефка Ангелова) caught the first one.
 - **Suite 6 may only compare a month with another month.** Every sheet in
   `test/generate_lifecycle.py` is internally correct on purpose — the arithmetic
   reconciles, the bases are right, each month would pass suites 1–4 alone. The only thing

skills/trz-expert/SKILL.md

@@ -253,6 +253,12 @@ fs = openpyxl.load_workbook(path, data_only=False)    # формулите
 (`references/proverki/k.md`) носи и „Как се чете електронна таблица“ пред своите
 проверки, и „Формули“ след тях.
 
+За I1, I5 (само липсващи дни болничен при начислена сума), B1, B4, B5 и F5 скилът носи
+`scripts/audit.py` — смята ги директно от файла, вместо да ги пресмяташ наум; ставките
+идват свежи от `references/stavki/` при всяко изпълнение, не са вградени в скрипта.
+Пусни го и вземи находките му за тези шест, преди да преминеш ръчно през тях; другите
+проверки на B, F и I остават на анализа тук.
+
 Изпълни проверките от всяка отворена група ред по ред. Всяка има изричен резултат — един от
 пет: преминава, не преминава, **недостатъчни данни** (проверката се дължи, но нещо липсва —
 назови точно какво), **непроверимо** (документът е подаден, но не става за проверка —