# Architecture · One row, and everyone allowed to touch it · receipt · 12 September 2026

Chapter one of _Shared systems_, delivered in the Eve edition on 12 September
2026. Twelve inherited slides become a thirty-four-slide chapter under the
`architecture` id. The brief is
[authoring/architecture-chapter.md](../../authoring/architecture-chapter.md).
Machine-readable state is in [state.json](state.json).

## What was delivered

Thirty-four slides in the guide: thirty-three authored here, plus
`map-of-the-platform` carried unchanged from Orientation. Twenty-two are new — a
cover, six section dividers, a close and fourteen teaching slides. All eleven of
the architecture guide's own inherited slides are rewritten in place; seven of
them belong to the reviewed foundation (`foundation-coverage.json`) and keep
their identifiers for that reason. Nothing is merged away.

This chapter opens a new track. `shared-systems` is declared in the catalog with
a reading route carrying this one chapter; the track's five later chapters join
the route as they are rewritten.

## Evidence

No capture. This chapter has no member-facing surface of its own — its subject is
a row, a registry and a route table — so every claim is executed or counted at
the pin instead.

- **The partial-failure envelope, executed here.** It refuses an envelope that
  claims to be partial with no errors, and one that claims to be whole while
  carrying errors. The builder derives the flag from the errors rather than
  trusting the caller, and throws on a contradicting one.
- **The contract registry and its schema, executed and parsed here.** Ninety-
  eight object contracts and thirty-nine enum contracts across twelve domains,
  ten control fields, one Prisma schema. That schema holds a hundred and
  forty-seven models: ninety-nine with the full control envelope, ninety-eight
  with the `(tenantId, contractSchema, sourceRecordId)` unique identity — the
  same ninety-eight the registry registers — and four hundred and ninety-one of
  four hundred and ninety-two indexes on those models leading with `tenantId`.
  The one exception is the tombstone audit table, indexed by domain.
- **One contract write, executed here.** Against a declared Prisma delegate
  double that filters on `where` the way the real delegate must: a Veritas
  timeline's six fields stored as sixteen columns; the payload hash identical
  under reordered keys and different under a changed value; a record its
  contract rejects refused; a tombstoned identity refused with the repository's
  own message; a tombstone matching one row, tombstoning one and writing one
  audit event while leaving the title and payload hash readable; and a purge
  removing the row.
- **The erasure cascade, executed here.** Every target attempted, `erased` /
  `not-found` / `failed` distinguished, `complete` false when one target failed
  and true when every target was already absent.
- **The OpenAPI spec builder, executed here.** Seven documents, three hundred
  and thirty paths and five hundred and ninety-nine schema definitions generated
  from the registry — and zero routes in the service matching any generated
  path.
- **The read-capability registry, executed here.** Eleven capabilities for the
  shell role, nine for the assistant, seven for an administrator — and eleven
  again when no role is supplied.
- **Counted at the pin.** 1,685 route registrations, of which 793 begin `/v1`
  and exactly one is mounted under the registry-declared prefix. 518 files using
  the authentication pre-handler factory and 510 the abuse one. 16 non-test
  files referencing the partial-failure envelope, ten of them member-facing
  routes. 106 non-test files importing the persistence library. 464 store
  modules, 74 of them durable. 37 migrations with zero column renames and zero
  column drops. 28 data-rights artifact kinds against 17 export collectors and
  16 deletion services.
- **An independent survey** of the shared architecture was run against the pin.
  Six of its findings were re-verified here before use, and one of them
  corrected a claim this chapter had already drafted:
  - the chapter's first draft said the same record under two tenants is two
    rows, which is what the compound unique key says. It is not what the
    database enforces: on ninety-one of the ninety-nine envelope models the
    primary key is a bare `id String @id` taken from the contract payload and
    not scoped by tenant, so the second tenant to write a natural key the first
    used is refused. A fixture checked into the repository measured exactly that
    against real PostgreSQL 16.14 on 2026-08-15, and records the write path as
    leaking while read, search, history and audit hold;
  - only five non-test files in the service read a tenant context off a request,
    and the one live writer through the contract service passes a fixed tenant —
    by design, because the editorial workbench it serves is a declared
    single-tenant surface;
  - residency fails closed on the way out, but the guard that would deny such a
    request before routing has no production caller;
  - the envelope's typed-schema factory has no production caller at all, and
    every route that uses the envelope is a read.

### What the measurement could not see

The delegate double keys rows by the tenant-scoped `where` clause, so it cannot
produce the primary-key collision between two tenants. That limit is recorded in
the evidence file beside the probe output, and the chapter reports the collision
from the schema and the checked-in fixture rather than from this run.

## Checks

| Check         | Result                                                              |
| ------------- | ------------------------------------------------------------------- |
| Build         | 1,613 slides, 711 sources; freshness check clean                     |
| Python suite  | 481 tests pass                                                       |
| Node suite    | see the commit body                                                  |
| Browser sweep | 34 slides × 1440, 1280 with audio chrome, 390 and print — 0 findings |
| PDF           | 34 pages, exact authored visual text, the one table complete         |
| Narration     | 1,613 tracks verified, nothing pending; 45.0 minutes in this chapter |
| Bindings      | no assignment or teaching file binds an architecture slide           |

The sweep passed on its first run.

## Treatment deviations

Three, each deliberate:

- **No capture.** The chapter's subject has no surface; the plan's "one real
  capture per chapter" is answered by executed evidence instead, and every
  number on every slide names the module it was produced by.
- **`architecture-read-handoffs` moves to `sequence-lanes`** as the plan
  directs, which retires the `ownership-handoffs` layout's only contract-test
  specimen. The layout is still live in four field-guide slides, so the five
  ownership tests were repointed at `sophia-live-flow` — the same shape, three
  owners and four ordered exchanges — and rewritten to derive their fixtures
  from the specimen rather than hard-code one slide's text.
- **One legacy diagram retired.** `partial-failure-contract` carried a
  `diagrams.json` graph entry that overrode its new card grid. The entry is
  removed, and the build test that used it as its no-declared-geometry specimen
  now uses `isis-dispatch`.

## What this chapter does not establish

Nothing here touches a deployed service, a real member account or a live
database. The contract repository was executed against a declared delegate
double, not Postgres. The cross-tenant collision the chapter reports was
measured by a fixture checked into the repository, not by this run. No claim is
made that any row reached a database on this host.
