fix(desktop): keep "return to latest" at the tail

[orangeCatDeveloper]

Summary

Fixes #4882

Clicking 返回最新消息 after paging far back flashed the tail and jumped back to the Turn the reader had been on; fast wheel paging occasionally stalled one page short. Together they are the transcript-scroll-cost.spec.ts:239 flake on main.

Root cause of the jump-back: ChatView pinned the scroller to the tail only after await onClick() resolved, but the tail batch reaches the store before that. In the gap the unpinned scroller reported the top Turn as the reading anchor, and the restoreRange effect (which re-runs on every messages change) found it missing from the new range and called loadAround on it. The fix pins first: a pinned scroller reports no anchor, so the range that arrives has nothing to be pulled back to.

Root cause of the stall: loadTranscriptHistory returned silently while a load was pending. The scroller asks on every reader movement, so the request behind an in-flight load was lost and nothing re-asked. The guard now lives in transcript-reading-position.ts as loadTranscriptHistory, keeps the last request that arrived mid-flight (latest outranks earlier) and replays it when the current load settles, still gated on the session and controller being current. Moving it out of app-shell.tsx is what the renderer architecture ratchet asks for and leaves the shell smaller.

The gate itself was one shell-wide object, shared by every Session. A navigation on the Session switched to next queued behind the first Session's in-flight load and was then dropped when that load settled under the old Session's currency check — the "Return to latest" click did nothing. The second commit keys the gate per controller (WeakMap<object, TranscriptHistoryGate>); each Session's in-flight load and queued request stay to themselves, with the same-Session queue semantics unchanged.

Overlaps with #4560, which rebuilds this area and removes the notice; that PR clears the anchor in app-shell.tsx before the load, which locally still loses the race (the scroller re-reports the anchor 17 ms after the load starts).

Verification

apps/desktop: npm run typecheck, renderer ratchet --write then --base $(git merge-base HEAD origin/main), biome check on the touched files, @maka/ui test:dist — all pass.

transcript-scroll-cost.spec.ts -g "paging back" --repeat-each=60 --workers=4, macOS, main 411512b vs this branch:

  3 failed
  17 passed (1.4m)          # main, --repeat-each=20
666735 [history] batch hasNewer=false hasOlder=true through=239 frags=20 evicted=220 idsMin=220 idsMax=239
666749 [history] restoreRange loadAround turn=turn-prompt-rail-1 seq=0
666811 [history] batch hasNewer=true hasOlder=false through=239 frags=20 evicted=220 idsMin=0 idsMax=19

  60 passed (3.1m)          # this branch, --repeat-each=60; 0 restoreRange loads after the tail across 60 runs

The instrumented lines are from a throwaway build of main with console.info in the store, the restore path and the shell; the fix branch is un-instrumented.

Second commit adds cross-session gate coverage to apps/desktop/src/main/__tests__/app-shell-session-ui-state.test.ts and re-runs the paging-back regression on that head:

  ✔ starts another Session history load without waiting behind an in-flight load elsewhere (0.224583ms)
  ✔ leaves the switched-to Session untouched when a stale Session load settles late (0.192958ms)
  ✔ reports a late history load failure to its own Session alone (0.114042ms)
  ✔ replays the queued latest load once after an in-flight load settles on the same Session (0.18725ms)
  ✔ keeps the queued latest load when an earlier request arrives after it (0.1795ms)
  ✔ does not replay a settled load after its Session range was replaced (0.16225ms)
  ✔ replays a queued load through its own Session callbacks (0.15575ms)
ℹ tests 27
ℹ pass 27
ℹ fail 0

  60 passed (3.4m)          # 76692daa8, --repeat-each=60

AI use

Select exactly one:

• No generative tool made a substantive contribution
• Generative tooling made a substantive contribution

Tool(s) and scope: Claude Code — trace analysis, instrumentation, root-cause bisection, the code change, and the per-controller gate follow-up with its regression tests; verified by the author.

Checklist

• Tests cover the change and fail without it
• Lint, format, typecheck and the affected suites pass locally

Does this PR entail a change in behavior?

• Yes — described under Summary above
• No

[hqhq1025]

Reviewed exact head 278beecd7ef0947ddbe266fe080703d59797ca25.

The patch fixes the same-session tail-pin race and replays a later same-session history request, but the shared replay gate loses session/controller ownership. The inline P2 below remains reachable during normal session switching, so this is not ready to merge.

Validation completed on Node 24.18.1: clean install, build:test, Desktop typecheck, Desktop dist tests (2211/2211), renderer build, architecture checks (101/101), Biome lint/format, locale hygiene, ASF headers, diff check, and the paging-back Electron scenario repeated 20 times (20/20). A deterministic helper-level probe reproduced the cross-session misrouting. The exact-head hosted test check and a local full Storybook smoke both fail only on the unrelated product-sidebar-session-list--project-groups focus assertion; that required check still needs to become green. I did not run the native macOS performance benchmark.

Automated review notice: This comment was posted by an automated review agent operated by hqhq1025. It is not an independent human review and does not replace one.

[hqhq1025]

Re-reviewed exact head 76692daa8550f5c6a5f641cd8b16dc082db064ec.

The follow-up keys history-load ownership by transcript controller, so a load for the newly active Session starts independently instead of being queued behind and discarded with the previous Session. The added regression cases also cover late settlement isolation and same-controller latest-request priority. The prior P2 is resolved, and I found no new P0-P3 issues in this incremental change.

Validation on Node 24.18.1: clean install, build:test, Desktop tests (2218/2218), focused session UI-state tests (27/27), Desktop typecheck, renderer build, architecture checks (101/101), Biome, locale hygiene, ASF headers, and diff check. The exact-head hosted test check was still in progress when this review was posted, so merge readiness remains gated on its terminal result and independent human review.

Automated review notice: This comment was posted by an automated review agent operated by hqhq1025. It is not an independent human review and does not replace one.

[hqhq1025]

Re-reviewed exact head f8603a025e993845e236c310b35852d82c875c3a.

The increment since 76692daa8550f5c6a5f641cd8b16dc082db064ec changes tests only. The production per-controller history gate is unchanged, and the prior cross-Session P2 remains resolved. I found no new P0-P2 issues, but the two inline P3 comments identify assertions that do not currently prove the behavior they claim.

Validation passed on Node 24.18.1: build:test, Desktop typecheck, UI tests (385/385), focused Desktop tests (27/27), renderer architecture checks (101/101), Biome, formatting, ASF headers, diff check, and a clean patch-preserving merge-tree against current main 2bfea3c32eb9eb3a907d1a5a04a8229156810b26. The required hosted test check is green.

Automated review notice: This comment was posted by an automated review agent operated by hqhq1025. It is not an independent human review and does not replace one.

[hqhq1025]

[P3] Observe the authority that ChatView actually uses. This probe is mounted as a sibling of ChatSurfaceLayout, under its own TranscriptScrollAuthorityProvider, while the layout creates a separate provider around the chat. As a result, view.authority starts and remains independently pinned, so the assertion at line 275 can pass even if the production pin-before-load ordering regresses. Mount the probe inside the layout-owned provider (or otherwise expose that exact authority) so the assertion observes the state under test.

[hqhq1025]

[P3] Do not use these queued-call promises as evidence that replay completed. loadTranscriptHistory returns immediately from the gate.pending branch, so both promises are already settled before the first load finishes; this await only happens to follow the replay continuation in today's microtask ordering. Either return a promise tied to the queued replay or await an explicit completion signal before asserting the final pending/callback state.

[Astro-Han]

Reviewed at f8603a0 against main (2bfea3c32e), +564/−29, seven files, mostly tests.

Both problems are real and both fixes sit at the owner. The jump-back is two lines: chat-view.tsx:747-752 now calls pinToTail() before await onClick(), so the reader anchor is cleared (use-chat-scroll.ts reports undefined while pinned, captureTranscriptReadingAnchor clears it, restoreSessionTranscriptRange has nothing to restore) before the range load resolves. That is inside the #4269 pin-owns-anchoring model, not a patch around it. The stall is the in-flight gate moving from the shell's boolean into loadTranscriptHistory (features/conversation/controller/transcript-reading-position.ts) and becoming queue-and-replay, keyed by controller in a WeakMap; since app-shell-effects.ts:415-448 builds one controller per active Session, the gate dies with its controller and the cross-Session P2 from the earlier review is closed. No new state mirrors the scroll authority.

Nothing blocking. Three P3s and two tests that can go:

• packages/ui/src/use-chat-scroll.ts:140-142 still documents the old contract ("it refuses a request while one is in flight"); this PR is the reason that sentence is now wrong. Change it to "queues the newest request and replays it".
• In return-to-latest-pin.test.tsx the AuthorityProbe reads a sibling TranscriptScrollAuthorityProvider, not the one chat-surface-layout.tsx:106 wraps the view in, and a never-attached authority starts pinned = true, so the pinned assertion passes on main too. The anchors assertion is the one with teeth (on main the third undefined never arrives). Drop the probe, the getter and the import.
• The second test in that file scrolls onto the position pinToTail() just wrote, which the authority's echo branch swallows before any reader decision, so it proves "own echo is not reported", not what its name says. Delete it. Likewise replays a queued load through its own Session callbacks in app-shell-session-ui-state.test.ts restates two neighbouring cases.
• app-shell.tsx:458 useRef(new WeakMap()) allocates a map per render; harmless, ??= on first use if you touch the line anyway.

Body: "leaves the shell smaller" is −7 tokens in the ledger; call it net flat, with the concurrency semantics moved to a helper that has tests. Also worth one sentence on why the gate lives beside the controller rather than in createDesktopTranscriptRangeController itself, which is where "one load per transcript, newest queued" naturally belongs; the answer is presumably that setPending / onError are shell concerns, and that is fine.

No file or semantic overlap with #4259; #4885 fixes the smoke flake that blocked this one's required check.

Evidence boundary: static read against main plus the authority, range controller and shell effects; the 60-run spec loop, the 27 desktop tests and the ledger number are the body's claims, not rerun here.

AI-assisted review: drafted with Maka; I verified the pin-before-await ordering, the per-Session controller lifetime and the probe's provider scope myself.