fix: B1/B5 read the month's norm from the day sum, not the worked-day column

[svedbg]

What was red

The Monday deep-suites run has been failing since 2026-09-07 — three scheduled runs, most recently 35595675501. audit_test.py --seeds 3000 reported 21 B1/B5 findings on rows generate_wide.py injected no minimum-wage scenario into. The 300-seed push job never sees it, which is why it sat.

Two different causes, and only one of them was a bug.

Eleven real false positives (ten B1, one B5)

B1/B5 only compare a row against the full МРЗ when it worked full time for the whole month — чл. 1, ал. 2 НСОРЗ prorates the floor for anything less. With no public-holiday calendar, "the whole month" was stood in for by the highest отработени дни on the sheet. That is only the month's norm if somebody that month had no leave and no sick day at all, and on eleven of 3000 seeds nobody did: the sheet's own maximum was itself short of the norm, and a row sitting exactly on it was compared against the undiminished floor with its own pay legitimately prorated below it.

seed 339, July: 21 worked days + 2 sick out of a 23-day norm, основна prorated to 584.42, reported as under the МРЗ.

The norm is now the largest отработени + отпуск + болничен + майчинство any row declares. Someone employed the whole month sums to the norm whatever mix of the four they had, so this is right as soon as one row was employed all month — a far weaker condition than one row with no absence at all. A sum is never below its own first term, so the new estimate is never below the old one: it can only silence a comparison, never open a new one.

A fractional value contributes nothing to that sum. It is an amount typed into a day column, which is exactly what K2 reports it as — counting seed 339's 161.02 sick "days" would read that month as a 181-day one and gate the whole sheet off instead. The two places that apply that rule now share DAY_TOL rather than each spelling out 0.005.

Ten that were not false positives

m_insurable_unexplained restates осигурителен доход at 86–95% of the correct figure. On a row already near the floor, the restated figure lands below МРЗ — a real second violation of the same understatement, not a false positive, and the same pairing the hand-built s_b5_insurable_below_min_wage shape already expects in the other direction (B5 and F1 together).

run_wide() now excuses exactly that and nothing else: B5, on a row carrying the injected F1_insurable_unexplained, with the F1 finding actually reported at that row. B1 anywhere, or B5 anywhere else, still fails like a miss.

Proof

Both halves are pinned by a new shape in audit_test.py part 3, each proved by sabotage:

sabotage	result
norm reverted to the worked-day column	s_b1_norm_only_visible_in_the_day_sum red, and seed 339 comes back
fractional exclusion dropped	s_k2_amount_does_not_inflate_the_norm red — a real B5 gated off the sheet

The МРЗ itself is written down in neither file, including in prose: scripts/rates.py reads it from the reference at call time and test/rates_test.py is the one place that pins it.

CLAUDE.md's note on this check's false-positive history gets the third round, including that only the Monday run can see this class of bug. One stale claim in the test fixture is corrected while there: the base rows were said to model "nobody has the true norm", which they never did — with no leave or sick day declared anywhere there is no evidence of a longer norm, and that row's основна is above МРЗ regardless, so nothing was being pinned. The new shape pins it for real.

Verified

• python test/audit_test.py --seeds 3000 — the failing step, green
• python test/run_tests.py --seeds 300, k_checker_test.py --seeds 300, komplekt_test.py, lifecycle_test.py, rates_test.py, skill_test.py, checks_test.py, preflight_test.py — all green

No change to SKILL.md's guidance, so no paid eval is implied by this.

🤖 Generated with Claude Code

https://claude.ai/code/session_01DpdxJWRY9P3YGknzhEzBw8