Card 78 — wired into every write surface; the watchdog is green again S4 DONE

The maths was done; today was plumbing — the step that once went catastrophically wrong. Computing a number inside the engine is not enough: it must survive the journey into the database. That journey crosses about ten hand-offs — the engine's in-memory record, two different columnar serializers (one for live streaming, one for the "dead letter" emergency spool that catches bars when the database is briefly unreachable), a Python dictionary bridge, the master column catalogue, the database blueprint, and the human-readable column documentation. Incident #556 happened because eight features were computed perfectly but five of those hand-offs still carried an old hand-written column list — the values were silently dropped at the boundary and the columns deployed 0% populated.

Today every hand-off learned about the new column, and — more importantly — every watchdog that exists to catch a missed hand-off was exercised and is green:

• The schema contract watchdog (red on purpose since iteration 02) re-checked that four independent descriptions of the table — the Rust record, the database blueprint, the feature catalogue, and the column documentation — now list exactly the same columns. It flipped green.
• The exact-count tripwires were deliberately moved from 85→86 and 90→91. They're exact equalities, not minimums — they fire on a missed addition AND an accidental removal, on both the Rust and Python sides.
• The round-trip rehearsal: Rust wrote a real 91-column file through the emergency-spool path and Python's data library read it back and verified every column name, order, and the new column's value — proving the value survives an actual serialize→disk→read journey, not just a code review.

Why this feature's number is trustworthy end-to-end now: the reference answers (iter 01), the bit-exact engine (iter 02), and the persistence plumbing (today) are each guarded by their own failing-by- default test. What remains is certification: formally proving the feature never peeks at future data (causality), always gives the same answer on the same input (determinism), surviving a full quality gate, and an adversarial review — then the pull request.