tri t27b: honest in-reference ratio, compile-only passes, frontend disagreements, per-spec delta between lab runs #6184

Refs #6063 (follow-up to #6112 / #6114)

Problem

On 2026-10-04 tri t27b status printed t27b 70 / reference 648, and the steward posted that to the epic. The honest figure is 56 of 648: 14 of the 70 passes are specs the reference itself fails. Two more facts in the lab's per-spec results show up in no summary number:

• 29 of the 56 in-reference passes have 0 tests and 0 invariants. They compiled and checked nothing.
• specs/port/scripts/gen_w378.t27: t27b's frontend rejects it (cannot assign I32 to ?i32) while the reference passes it. That is a disagreement, and mismatch does not count it.

None of it is visible between two runs either: the lab keeps runs/<sha>.json, but nothing compares them per spec.

Change

• tri t27b status prints the in-reference ratio, how many of those passes ran a check and how many are compile-only, and the out-of-reference passes separately.
• tri t27b doctor gains LAB-FRONTEND-DISAGREES.
• New tri t27b delta [--from SHA] [--to SHA], a per-spec ratchet over two lab runs: REGRESSED, NEW-MISMATCH, CHECK-LOST (exit 1), GAINED and denominator moves (reported).
• Fixture tests extend test_a_t27b_tick_reads_before_it_acts.py.