# Bounded privacy-outcome teaching execution

Use the unchanged TEACHING-ERASURE-1 revision1 policy. Its fictional product
owner is assumed to have approved it ONLY inside this synthetic scenario.
Eligibility, identity, applicable exceptions and scope were resolved outside
the fixture. No actual customer approval, legal applicability decision or
compliance result is asserted or inferred from a database query.

The tested boundary is the exact retained Python `buggy_delete(stores,subject)`
and `Request.attempt()` / `Request.reconcile()` over exactly two named SQLite
live stores, profiles and analytics. Each starts with the literal target and
unrelated control subject/value rows in expected.json. The target is the same
eligible synthetic subject in all calls. There is no type/parser/authentication
adapter, native privacy connector, network or external recipient in this test.

The original runner and its9 count-based checks/6 scenario observations are
preserved and rerun once into a new report. Their original Oct3 report is not
overwritten. A new literal oracle separately checks complete ordered
(subject,value) rows, exact response/snapshot keys and recursive value types,
and association of destination name with its actual observed outcome. It does
not trust Request.state, Store.contains() or an aggregate flag as its state
oracle. SQL observations come directly from each named live connection.
Expectations are authored from the unchanged fictional policy and declared
fixture inputs, not copied from the implementation's returned result.

The first bounded gate calls the actual buggy handler with failure injected
before analytics DELETE. A profiles success and analytics target row remaining
cannot satisfy the completion requirement. The independent expected state is
unresolved. Actual complete therefore FAILS this gate; its expected rejection
must not be reported as buggy correctness.

The corrected gate calls the actual Request methods for five sequential
observations: failed analytics attempt, recovery retry, completed-request
repeat, explicit reconciliation after an actual analytics backup restore,
then deletion of the restored live row. Literal snapshots expect attempts
1,2,3,3,4 and exact named statuses. The attempts field is observable API metadata,
not a universal policy rule: repeating completion changes2→3 intentionally.
Idempotent logical effects here mean the two live row sets and completion/
absence semantics stay stable on repeat, not that every snapshot byte is equal.

The backup is taken before deletion using SQLite's actual backup API. It is
restored into analytics, and only an explicit Request.reconcile() call reopens
the request. Cached complete state before reconciliation is retained as a
diagnostic observation, not a claim of automatic restore detection. The backup
still has both original rows before cleanup, including the target. No backup
erasure is performed. Removing the temporary directory is ordinary test
cleanup, not evidence of physical sanitization or erasure of every copy.

One designed report-adapter control calls the real corrected partial-failure
method, then swaps only the profiles/analytics status values in its copied
snapshot. Aggregate state remains unresolved, while destination attribution
is wrong. The literal full snapshot oracle must reject it even though an
aggregate-only state comparison passes. This is an injected wrong report
mapping, not a discovered product defect, another network event or a new policy.

The wrapper runs four child gates: original replay=0, buggy outcome=1,
corrected sequence=0, swapped report=1. Wrapper success means these expected
contrasts were observed; it does not accept the buggy handler or wrong adapter.
Fresh original9 checks and six observations remain separate from the new
seven observations (one baseline, five corrected, one adapter control).
Capture actual commands, times, versions, source hashes, call identities,
per-store SQL channels, complete rows and response differences.

SQLite files exist in a temporary directory, but all observations use the same
live connections. Record actual connection autocommit/isolation/transaction
settings; the unchanged Store source does not pin modern connection defaults.
The injected DestinationUnavailable occurs before a DELETE and leaves the
connection readable by the oracle. No post-write SQL failure, outage behavior,
reopen/restart durability, power-loss recovery, concurrent writes, automatic
restore monitor, durable request queue/ledger, downstream inventory, recipient
notification, deadline process or physical media erasure was exercised.
