[B] #514 slice 6: regenerate the coupling matrix from the ledger, and delete the hand-written one

Slice 6 of #514 (modular monolith). Track B — enforcement only, no production code. Needs #842, #845, #846.

Re-derive the design's §3.4 coupling matrix from a walk, and delete the hand-written one.

Why

The 2026-08 matrix was assembled by hand and was wrong in five cells within five weeks — three
genuinely new edges it marked —, and two artifacts it would have counted as edges:

Cell	Verdict	Evidence
Farm → Commerce	real, `R`	`UpdateFarmSettingsHandler.cs:12`, `:149` (#727)
Access → Commerce	real, `R`	`SetStepperUnitHandler.cs:13`
Commerce → Access	real, `R`	`ConfirmSaleHandler.cs:23`, `:171` (#727)
Customers → Sales	artifact	`Customer.cs` lives in `src/Cluckwork.Domain/Sales/` — both ends Commerce
Accounts → Media	artifact	`Domain.Media` has no entity and no table

A matrix maintained by hand is a matrix that is wrong five weeks later. This slice makes it output.

Scope

Classify every edge the [A] #514 slice 1: module ledger and cross-owner edge ratchet #842 ratchet knows about by kind — W write, R read/validation,
E after-commit event, Q read-model query, P platform service.
Generate the matrix from the ledger plus the table-ownership walk ([B] #514 slice 3: table-owner completeness, walked from the EF model #845).
Assert the generated census and the ratchet's edge set are identical, so the two cannot drift.
Replace the design's §3.4 table with the generated one and say where it comes from.

Done when

Mutation	Expected
add an edge to the ledger without a `kind`	red
hand-edit the generated matrix	red — regeneration disagrees
census and ratchet edge sets differ	red

Explicitly not in this slice

Does not decide whether any edge is correct — only that it is recorded and classified. Removing an
edge is Track C.

test(arch): build the matrix diff header outside the StringBuilder co…