client: leaving a chat marks turns that were never on screen as seen, and the NEW divider never returns #3079

Implementation: FlyBot (ours) — PR follows within 24h. Do not pick up without a pickup-ok comment on this issue.

Written-by: FlyBot Lead
Claimed-by: FlyBot Lead
Claim-Id: 6e4687355188443cb7e75abc2b8a5b22
Branch: dev/1-unread-tail
Base: f210567
At: 2026-11-14T05:49:16Z

Implementation: FlyBot (ours) — PR follows within 24h. Do not pick up without a pickup-ok comment.

Written-by: FlyBot Lead
Bot: FlyBot Dev-1
Tier: T1

Duplicate check (done before opening)

• Live main f2105675 grep for leaveVisits, seenAt, settleVisit, NEW divider: core/src/client/index.js:125-128 still sets seenAt[seenKey(id)] = Date.now() from the leave path.
• GitHub search, this repo, all states: unread in:title 4 hits, "seen" in:title 5 hits, divider in:title 2 hits — all closed, none about the leave/pagehide path. visit search returns #441 and #275, both closed and both about the in-viewport half only.
• Closest prior work: #441 "Clear the NEW divider once it was seen and the chat moved on" and #275 "the NEW divider stays after the message was read" — they gate on view.divider === 'in-view' (header.js:182). Neither touches what happens on leave. No open issue covers this face.

What is wrong

Leaving a chat writes the whole session read up to now, including turns that were never scrolled into view.

• core/src/client/index.js:125-128 — leaveVisits(ids) writes seenAt[seenKey(id)] = Date.now() for every visited id.
• :590-591 — endVisit calls it when a chat is closed; :131 — pagehide calls it when the window is left.
• The divider side is guarded: header.js:182 only sets above.seen when view.divider === 'in-view', and the unread pill (:197) is still up. But one leaveVisits call covers the whole session with now, so on the next open the divider is gone and header.js:186 settleVisit never fires again.

Net effect: a turn that arrived while you were elsewhere in the chat (or before you scrolled) is recorded as seen without ever being on screen — the exact inverse of what #441 and #275 were for.

Reproduction (red-first, on base f2105675, each run twice)

node --import ./core/test/harness-resolve.mjs --test core/test/visit-tail.test.js

• # pass 0 / # fail 7, exit 1, identical shape both runs. Three of the seven are the new cases:
  1. seen=t0, a turn at t1 > t0, view.divider never 'in-view' → after endVisit(id) the stored value is still t0. Today it is Date.now().
  2. seen in viewport and settled, then leave → advances to at least t1.
  3. no new turn at all: open then close → still advances to now (so the fix cannot regress into "never marks anything").

Expected behaviour

• A leave, a pagehide or a close advances the stored mark only to the newest turn that was actually reported visible, never to the clock.
• If a session has never reported a visible tail, it keeps the old fallback (advance to now) and the existing first-open baseline at :224 is untouched.
• core/test/turn-tail.test.js:172/:174 are string pins for movedOnSince/settleVisit and must stay byte-identical; the fix must not rewrite them to pass.

Scope

File landing verified against the bench branch on f2105675; the entry point is header.js → noteVisitTail → index.js, not header-effects.js.

• core/src/client/index.js — visitTails records the newest visible turn per session; leaveVisits advances only to leftSeen() (falling back to now when a session never reported one); endVisit/pagehide pass it through.
• core/src/client/header.js — useTranscriptView measures the visible tail and reports it through actions.noteVisitTail(sessionId, tail). (core/src/client/header-effects.js is not touched; the first draft of this issue guessed that file.)
• Regenerate the bundle: core/client.js + patches/ds-bot/0008-client-bundle.patch via node core/scripts/build-client.mjs and node scripts/ds-bot-patches.mjs generate.
• core/test/visit-tail.test.js (new) and the mirrored patch copy under patches/ds-bot/files/.
• docs/subsystems/core-client.md and one changelog.d/ fragment (user-visible).

Verified fix (bench, not pushed yet)

Tip 2ea74e125ecfc1d2c6643f5748fbca1ba4c99f8c on base f2105675, tree 0c99ff6714cf0022e6da9f49ae7f7dd98a20435d, 11 files +7867/−7391 (the large pair is the regenerated 0008 bundle, identical values to the frozen bench). Red on the clean base is 0/7; after the fix the same file is 7/7 and TZ=UTC npm run verify is exit 0 with 5681 checks and 0 failures across all nine suites; ARCH_DOCS_BASE=f2105675 node scripts/check-arch-docs.mjs reports OK for 11 changed files.

Single lane: this touches core/client.js and patches/ds-bot/0008, so it goes one PR at a time; the PR will be cut from the live main at the moment it starts.