[Amperstrand]

Draft — raised to showcase a problem, not to propose a final design.
The cost profile, measurements, and discussion live in #2631. This
branch exists so the maintainers can see one possible batching
shape against the real code — treat it as context for exploring the
problem, not a change ready for review. Closing it is fine if the
conclusion is a different shape.

What this sketch shows

A pre-flight inside recover_incomplete_sagas that:

• collects Send/TokenCreated candidates and their reserved proofs,
• asks the mint about all of them in chunked batch NUT-07 requests
  (256 proofs per request in the sketch — one round-trip where the
  sequential loop would make 256),
• for each saga whose proofs are ALL spent, applies the same
  spent-branch bookkeeping recover_or_complete_send applies (proofs
  Spent, transaction Pending → Completed, row deleted),
• defers everything else — unspent rows (revoke window), other saga
  kinds, and the whole chunk on any request failure — untouched to the
  existing sequential resume paths, so offline/flaky mints see no
  behavior change.

Deliberately open questions (why this is a draft)

• Where this belongs: inside recovery at all, vs. a runtime/periodic
  pass, vs. both. Our production daemon chose the periodic pass; this
  sketch puts it in recovery because that's where every wallet pays the
  cost today.
• Chunk sizing: 256 proofs/request is a placeholder.
• Interaction with refactor(wallet)!: unify resumable Rust and FFI workflows #2479: that refactor reshapes the recovery loop
  (per-saga locking, re-fetch). This pre-flight is structurally
  independent of the loop's shape, but would need to be re-expressed on
  top of it — hence sketching now rather than proposing.
• The SQL side: get_incomplete_sagas still returns every row; a
  state/kind filter might be the better or additional fix. We did not
  touch it here because that changes the primitive's contract for other
  callers.

Evidence the shape is viable

• cargo test -p cdk --lib: 730 passed / 0 failed (3 new tests: batch
  finalization, unspent rows retained, unreachable-mint deferral).
• One pre-existing test switched to a new additive mock helper
  (MockMintConnector::set_check_state_response_persistent) because the
  one-shot staged response is consumed by the batch call while the
  sequential path legitimately asks again.
• Clippy clean.

Feedback very welcome — especially "wrong place, here's the right one".