[joeychilson]

The status's historyReadAt was a time nothing reads, rewritten at every catch-up: serve's push() turned it into a bool before comparing, and each catch-up wrote the database for it. It's now historyRead, true once the first read of the agents' history has finished, written once. Protocol 2.

Flags: changes the protocol (1 → 2); edits 3 tests for the field's name and protocol number; changes what state.history_at holds (no new revision)

Changes

• historyRead replaces historyReadAt: whether state.history_at is set.
• ingest sets history_at once, when the first read ends, even with nothing to read.
• push() compares statuses less at, rewriting nothing.
• The app's reading: a status whose historyRead is false.

Evidence

contract/status.notification.json:

     "at": 1791460800000,
-    "historyReadAt": 1791460740000,
+    "historyRead": true,

New serve::tests::a_status_is_pushed_only_when_what_it_shows_changed: pushes false, then true after the first read, nothing after a catch-up that read a session but changed nothing shown, then one after a settings change.

New ingest::tests::a_catch_up_after_the_first_read_that_reads_nothing_writes_nothing fails on main: another connection's data_version moves (left: 3, right: 2).

The Swift engine test passes. Main's binary, this one, then main's again share a scratch database fine; rollback is the previous build (details).

Not checked: the app by hand; real user databases.

Risk

• A hello during the first read now gets historyRead: false; main set the time once files were listed, before reading them.

Left out

• db::REVISIONS[0]'s comment still says "last read"; shipped scripts aren't edited. architecture.md has the new meaning.

Review focus

• src/ingest.rs ingest: when history_at is written. The rest is mechanical.

Runs

Swift against the real engine:

$ cargo build
$ TURNSCOPE_BINARY="$PWD/target/debug/turnscope" swift test --package-path macos/TurnscopeKit -Xswiftc -warnings-as-errors
✔ Test theRealEngineIsStartedSaysHelloAndTakesChanges() passed after 3.232 seconds.
✔ Test run with 46 tests in 0 suites passed after 3.232 seconds.

Main's binary (old) and this branch's (new) on one scratch data directory, with TURNSCOPE_OFFLINE=1 and the fixture home's copy, target/tmp/home. Each sessions catches up, and each hello is followed by shutdown. jq prints a field that's missing as null.

fresh db, old binary: sessions
  history_at=1792449407450
old serve hello:
{"at":1792449360000,"historyReadAt":1792449407509,"historyRead":null}
new binary: sessions, twice (nothing new)
  history_at=1792449407509 then 1792449407509
new serve hello:
{"at":1792449360000,"historyReadAt":null,"historyRead":true}
back to the old binary: sessions
  ok, history_at=1792449407854
fresh db, new binary: before any read
  history_at=no db
  sessions twice: history_at=1792449407912 then 1792449407912

The push test fails at assert!(pushed(&mut server).is_empty()) when push() skips its comparison. It fails with left: [], right: [true] when the first read is not marked because there was nothing to read.

These were also run and passed: cargo fmt --check, cargo clippy --all-targets --locked -- -D warnings, cargo test --locked (again with TZ=America/Chicago), RUSTDOCFLAGS="-D warnings" cargo doc --no-deps --locked, shellcheck scripts/*.sh, actionlint, cargo deny check advisories, scripts/build-app.sh, and scripts/check-app.sh on that build. check-app.sh reported "serve: protocol 2, a hello answered with a status" and drew the panel from the contract's example.