[kvlonge]

Why

When a Recall session closed, its final delivery-summary rows were capped at 500 ms even when most of shutdown_timeout remained. A single slow hosted-store response therefore marked a fully delivered transcript incomplete. CI saw this on Windows: every stream was delivered, yet the session reported "Recall delivery deadline expired".

Changes

• The final delivery rows now use whatever remains of the shutdown budget, with a 500 ms floor. close still returns by the shutdown deadline, so shutdown_timeout still caps the delay at exit, as the docs describe.
• Because of the floor, rows that finish within 500 ms of the deadline can land after close has returned and reported history incomplete. The store, which is the source of truth, is then correct and the CLI under-reports, which is the safe direction.
• Two writer tests use a hosted store whose final rows are slow:
  • final rows that take 1.5 s inside a 5 s budget record complete;
  • final rows slower than a 1 s budget still close within it and report incomplete.
  • Both cases assert the deadline the store receives for the final rows is within the shutdown budget.

Verification

• The first new case fails on the previous code, matching the CI failure, and passes with the fix.
• With the final rows given a 30 s deadline instead, the tests fail.
• cargo test -p sqlbuild-recall: 67 passed.
• The orchestration test_reads test passes against PostgreSQL.
• clippy, fmt, fensu and make check-ci pass.
• DuckShop edge case: not applicable, because this changes Recall delivery timing, not a project shape, data value or lifecycle sequence.