bot: stop the redundant join and full replay after a table closes

[svanlabs]

What generated this?

Generated-by: claude-code/claude-sonnet-5-5

Motivation

Closes #967.

After a table closes the server re-seats the bot within 0.1 to 0.2 s, but the join that table_closed queued stayed set, because Seat::joined cleared only the leave. The session loop sent join_lobby about 10 s later and drew already_seated. The handler then asked for a resync from sequence 0 although the tracker already held the table, so the reply replayed the table's whole history over live state.

Live evidence on build 7a1494c (2026-10-10 11:00 to 2026-10-11 22:12 UTC), from artifacts/logs/svanbot10.log:

• Four table closes; three drew already_seated (Svanism twice, Svanar once) and one already_in_lobby.
• 37 false warnings "a replayed result for hand X cannot be settled: its start was not saved". Their number equals the hands played at the table before the bot sat down (16 at table 49a1713a, 3 at table edd14b44), and none of the 37 hand ids is in the history store. played_seat matches a result by seat number, so an earlier occupant of the seat counts as the bot.
• One stored hand carries state from replayed earlier hands: Svanism hand e1b37e33 records seat 2 folding preflop with bet_before 744 against a 30-chip pot. 1 of 13,486 hands since the install.

No row was lost and no action was illegal; the cost is false alarms on the one warning that exists to catch a lost row, and one polluted summary.

Solution

Three commits in crates/apps/bot/src/client/, two for the defects and one pinning a call site the review found loose. Each defect has a test that failed before its fix, run on the unmodified code first:

1. Seat::joined also clears pending_join. Test: a_seat_the_server_gives_us_makes_a_queued_join_moot (failed with "seated: nothing is left to join").
2. The already_seated handler goes through new recover::adopt_table: a table the tracker holds asks from its watermark, as a reconnect does; any other table still asks from 0. Test: an_already_seated_for_the_table_we_hold_asks_from_the_watermark (failed with left 0, right 40; both halves run against a tracker that holds another table, so an implementation that ignored the table id would fail). The protocol spec row says so. handler.rs goes from 500 to 499 lines.
3. A handler-level test that table_closed then table_joined leaves no join queued, which pins the table_joined arm that calls Seat::joined (it fails with the clearing line removed), and the two spec rows that said a close rejoins the lobby now say "unless table_joined arrives first".

Off the decision path: client lifecycle, nothing the golden snapshot or sim paired exercises, so no paired simulation. Not released: the running build is 7a1494c.

Not changed, filed separately: the stale resumable entry behind three other false "row is lost" warnings (#968), and the stale history store (#969).

Checklist

• scripts/check.sh full passes on this exact commit (e4e3df2), run in a worktree on the shared warm target/dev:

test.py: 896 passed, 0 failed, 4 ignored in 35 binaries; build 0.3s, run 4.5s (6 jobs x 2 threads, profile gate, peak test RSS 196 MB)
== ok (full, 144 s)
check.sh full: exit 0

• This is one change: the redundant join after a table close and the replay it triggers. Three commits: two defects in one area, and one test that pins a call site.
• Not a behaviour change on the decision path: no paired simulation; the golden snapshot is untouched.
• Not a performance change.
• The document that names the changed behaviour is updated (docs/SPEC-protocol.md, three rows).
• Generated-by: is in every commit footer and in the section above.

🤖 Generated with Claude Code