[Kiran01bm]

The demo tour now walks a data fault through the built binary: a copy-and-swap whose rows will not fit the new type reports failed / data-fault with exit 1, and a second copy-and-swap on the same table proves the abandon left nothing behind.

Why

The engine abandons a run on a data fault — drops the shadow, releases the slot or the fence, deletes the checkpoint row — and the Go suite proves that against testcontainers. The artifact smoke test did not exercise the path at all, so a regression in the CLI's rendering of it (the exit code, the code field, the table) would reach a release unseen. The tour is the one check that runs the shipped binary the way a user does.

While wiring it in, the seed turned out not to be rerunnable once a copy-and-swap has run: the compose role is pgsprite, the engine creates a schema of the same name for its checkpoint table, and from then on the seed's unqualified CREATE TABLE landed in pgsprite instead of public under the role's default search_path, so every public.* assertion in a second make demo-check failed. CI never saw it because its database is fresh.

What

• demo/seed.sql: orders gains a note text column; one row (id = 4000, past the copier's first chunks) carries a 64-character note. The seed pins search_path = public so reseeding lands in public whether or not the engine's schema exists yet.
• demo/tour.sh: new execute_data_fault helper (check mode asserts exit 1, outcome == failed, code == data-fault, table, not forced; demo mode prints the verdict and the expected exit) and two run_exec rows after the existing copy-and-swap: narrow note to varchar(32) (the fault), then widen it to varchar(200) through copy-and-swap again — the success is the proof that no shadow, slot, or checkpoint row survived the abandon.
• demo/README.md: the exec row and the check-mode exit-code list name the new step.

Before / after

Same table, same seeded rows. Before, the tour ran one copy-and-swap on orders and moved on; a data fault was never exercised by the built binary, and the second local run seeded into the wrong schema. After, the fault is met part-way through the copy, the verdict is asserted, and the next copy-and-swap on the same table succeeds only because the abandon released everything.

Before                                        After
──────                                        ─────
make demo-check (fresh DB)                    make demo-check (fresh DB, or a reused one)
  seed → public.orders(user_id int, total)      seed → SET search_path = public
                                                       public.orders(user_id, total, note)
                                                       id 4000: note = 'xxxx…' (64 chars)
  exec:                                         exec:
    orders.user_id int → bigint                   orders.user_id int → bigint
      executed-via-copy-and-swap, exit 0            executed-via-copy-and-swap, exit 0
                                                  orders.note text → varchar(32)
    (data fault: never run by the tour)             chunk [2001,4000] hits id 4000
                                                    → SQLSTATE 22001
                                                    → run abandons itself:
                                                      shadow dropped, slot dropped,
                                                      checkpoint row deleted
                                                    verdict: failed / data-fault
                                                    table public.orders, exit 1   ◀ asserted
                                                  orders.note text → varchar(200)
                                                    executed-via-copy-and-swap, exit 0
                                                    ◀ could not start if anything
                                                      had been left behind
    DROP INDEX orders_total_idx rows …            DROP INDEX orders_total_idx rows … (unchanged)

make demo-check again (same DB)               make demo-check again (same DB)
  pgsprite schema now exists                    pgsprite schema now exists
  seed → pgsprite.orders  ✗                     seed → public.orders  ✓
  every public.orders assertion fails           all checks pass

🤖 Generated with Amp (Claude Opus 4.6)