[medazizktata25]

Related Issue

None filed.

What

e2e_clone_robustness::clone_range_and_temporal_hydration_elide fails on every run that can reach it:

temporal window hydrated (t0=0 t1=0)

The test is out of date, not the extension. Its temporal half was written for a synchronous fetch: it expects the query on a date window to raise rows_fetched. The commit after it (f7a4d83) changed that path, and the test was never updated:

• an uncovered temporal window now federates to the source for an immediate answer;
• it enqueues a background time copy of the window;
• the worker's copies skip gfs.clone_stats (DEFER_BOOKKEEPING), so rows_fetched cannot move.

The data is right: the query returns exactly the source's rows. CI did not catch this because it runs none of this file's tests (they need the gfs-postgres image).

How

1. test(clone): the temporal-window case follows the async copy it exercises. The temporal half now checks what the path does:
   • the first query federates once (federate_calls +1) and returns exactly the source's rows;
   • the queued copy lands: nothing left in gfs.copy_queue for the table, and the table is whole-cached or holds a cached range (waits up to 60 s);
   • the repeat query returns the source's rows without federating again.
2. docs(clone): rows_fetched does not count the background worker's copies. A comment where clone_stats.rows_fetched is defined, saying what it counts: rows a backend copies for its own query, plus gfs.warm's, and not the worker's. A matching note in tpch_validate.sh, whose pulled() reads that column as rows pulled.

Review Guide

1. crates/applications/cli/tests/e2e_clone_robustness.rs: the temporal half of clone_range_and_temporal_hydration_elide.
2. crates/extensions/gfs/src/sql/schema.sql and tpch_validate.sh: comments only.

Testing

On Rust 1.93.1, macOS with Colima for Docker, and a gfs-postgres:16 image built from this tree.

• e2e_clone_robustness: 5/5 pass, and they ran (not the early return for a missing image).
• Calibration:
  • with an extension image whose worker never runs its typed copy jobs, the test fails with "the background copy of the window never landed";
  • with the image rebuilt from clean source, 5/5 pass.
• cargo fmt --check clean.
• Commit 2 changes only SQL -- comments and a shell comment, so the installed catalog is unchanged; the image was not rebuilt for it.
• Not tested: Docker Desktop, podman, a Linux Docker host.

Documentation

• The rows_fetched comment in schema.sql, and the note in tpch_validate.sh.

User Impact

None at runtime: one test and two comments.

Out of scope, found while checking this: the copy worker still shares locks with foreground queries, and a user query on a table being copied can fail with deadlock detected. I reproduced it; it will be reported separately.

• gfs.take_token() takes gfs.budget FOR UPDATE before its unlimited early return, and the worker holds that for the whole copy.
• The worker then updates the table's clone_source row, which a concurrent query holds from bump_access.

Checklist

• Self-review completed
• Format check passes (cargo fmt --check)
• This PR can be safely reverted.