[thekevinbot]

Fixes #1246

Status

Test-only change in packages/rust/tests/sdk.rs. just preflight passes (0 failing, 3 packaging skips). No e2e receipt or changelog fragment owed.

The race

fs::write opens with O_TRUNC and then writes. The watcher's per-file hook runs on the first inotify event, which can land between the truncate and the write, so read_to_string returns "" and the first Update (or Insert) carries an empty name. it_streams_watch_events had the same shape on new_item.txt; async_sdk.rs::it_streams_watch_events already avoided it by staging and renaming.

What changed

A write_watched helper in sdk.rs writes to a .tmp sibling (outside the **/*.txt glob) and fs::renames it over the target, so the watcher sees one atomic replace with the final bytes. Both it_streams_watch_events and it_streams_watch_update_events use it. No sleeps, timeouts, retries, or loosened assertions. The other writes in the file happen before watch() or feed an extractor that ignores content, so they are not exposed.

Reproduction

Serial runs never tripped it: 30 sequential runs of the prebuilt sdk test binary (filtered to the two watch tests) under stress --cpu 32 --io 8 --hdd 4 passed 30/30 (/tmp/dirsql-1246-repro.sh).

Concurrent copies did. Same binary, 10 copies at once per round under stress --cpu 32 --io 8, on a 32-core box (/tmp/dirsql-1246-rounds.sh <bin> <label> 10 <rounds>):

	runs	failures
before (origin/main)	30 (3 rounds x 10)	1, it_streams_watch_update_events: left: Text(""), right: Text("final") at sdk.rs:426
after (this branch)	100 (10 rounds x 10)	0

A first attempt at 30 concurrent copies produced 8 failures, all Too many open files (os error 24) from the inotify instance cap (max_user_instances=128, ~100 already in use); those are not the race and were discarded.

Manual run: cargo test -p dirsql --test sdk passes 37/37 on this branch.