perf(gui): share one timer per interval and skip redundant revisit fetches

devlog/_plan/260816_gui_loading_performance/040_phase4_poll_consolidation.md

@@ -160,3 +160,46 @@ else add cases to client-resource-revalidate.test.tsx):
 
 Server-side endpoint merging, render-level memoization audits, virtualization
 tuning, bundle-size work.
+
+## D addendum — landed (2026-08-16/17)
+
+Implementation: commit d6ea35ef0.
+
+### Measurement (CDP, sandboxed instance on :5198, same protocol as 001/E0)
+
+| Scenario | Before | After |
+|---|---|---|
+| Dashboard, 31s dwell, visible | 57 requests / 9 concurrent 5s timers | 57 requests / **1** shared 5s timer |
+| Integrations revisit inside 60s | full refetch of 8 overview resources, skeletons on mount | **0 requests, 0 skeletons** |
+| Hidden tab (any page) | WP3 guarantee | unchanged: zero timers, zero requests |
+
+The visible-tab request count is deliberately unchanged: bucketing removes wakeups,
+not cadence, and freshness (not volume) is what a live dashboard is for. The real
+volume win is the revisit path — previously every tab hop re-fetched everything
+because `scheduleStoreEviction` drops the store on route change.
+
+### Interval tuning decision (§3)
+
+No cadence changed. The measurement does not show Logs (2s) or Debug (1s+2s)
+dominating: they are page-gated (`enabled: tab === "logs"` / `active`), so they
+contribute nothing unless the user is looking at them, and both are now
+visibility-paused and in-flight-guarded by WP3. Changing a freshness-affecting
+default without the numbers supporting it would be exactly the kind of unforced
+regression this section exists to prevent.
+
+### Deviations from spec
+
+- The spec assumed the Integrations pages would gain hand-written cache wiring.
+  Ten resources across two files needed the identical seed+write pair, so
+  `useDataSurface` gained an opt-in `sessionCacheKey` instead and the pages pass a
+  key. Same mechanism, one implementation.
+- `startVisibilityPoll` schedules through `window.setInterval` when available. The
+  migrated pollers all used the window timer and their tests intercept it there;
+  the bare global bound to a different scope and broke nine tests (found by running
+  the full suite, fixed before commit).
+
+### Verification
+
+- `cd gui && bun test tests` → **922 pass / 0 fail** (157 files), exit 0.
+- `bun run lint`, `bun run lint:i18n`, `bun run build` → all green.
+- Browser: revisit flow measured above; dashboard poll wave unchanged and healthy.

devlog/_plan/260816_gui_loading_performance/050_delivery_record.md

@@ -0,0 +1,79 @@
+# 050 — WP5: delivery record (stack publication, CI, remote suite, merge)
+
+## Stack shape
+
+Four dependency-ordered layers, each PR based on the one below, bottom targeting
+`dev`:
+
+| # | PR | Branch | Layer |
+|---|----|--------|-------|
+| 1 | #1854 | codex/gui-resource-deadline | resource deadline |
+| 2 | #1855 | codex/gui-auth-unwedge | 401 re-bootstrap unwedge |
+| 3 | #1856 | codex/gui-hidden-pause | hidden tab = zero timers |
+| 4 | #1857 | codex/gui-poll-consolidation | shared scheduler + revisit freshness |
+
+## Rebase history
+
+Two cascades, both `git rebase --update-refs --onto origin/dev <old-base>` from
+the top branch so all four refs move in one pass:
+
+1. The stack was cut at `b81314cd2`; `origin/dev` had advanced 88 commits.
+2. A second cascade onto `8f7a22ff7` picked up PR #1853, which had landed the
+   admission-`source` contract. That one mattered: the remote suite had reported
+   2 failures in `tests/loopback-listener-admission.test.ts` that were NOT ours —
+   the remote checkout paired #1853's NEW test file with our older source. After
+   the cascade the same suite ran clean, which is what confirmed the diagnosis.
+
+A comment-text-only conflict in `gui/src/visibility-poll.ts` was resolved in
+favor of the WP3 wording (see below).
+
+## Layer independence correction
+
+CI caught what local runs could not: WP3 (#1856) failed its own `gates` job with
+nine failures in the codex-auto-switch and account-picker suites. Those tests
+intercept `window.setInterval`, and `startVisibilityPoll` was scheduling through
+the bare global — the fix existed only in the WP4 commit. Since DEV-STACK-03
+requires every layer to be green at its own tip, the fix moved DOWN into WP3
+(`116bcae03`, later `c75f88d6c`) where the migration itself lives.
+
+Verified after the move: WP3 tip 913 pass / 0 fail, WP4 tip 924 pass / 0 fail.
+
+## Verification evidence
+
+- Remote full suite (`ssh lidge`, `bun test --isolate tests`) on the WP4 tip
+  after the second cascade: **12700 pass / 0 fail / 15 skip, EXIT=0**.
+  A `dev`-baseline run in the same session also finished EXIT=0, which is how the
+  earlier 2 failures were attributed to the stale base rather than to this work.
+- Local: `bun run typecheck` green; `cd gui && bun test tests` 924 pass;
+  `bun run lint`, `bun run lint:i18n`, `bun run build` green.
+- Browser: dashboard renders unchanged; revisit inside the freshness window
+  issues zero requests with zero skeletons.
+
+## CI flakes encountered (not regressions)
+
+The macOS and `test 2/4` jobs each failed once with a **Bun runtime crash**
+(exit 133 / exit 132, `storage-worker-lifecycle` under singleton isolation) —
+the workflow itself distinguishes these from assertion failures and retries once
+before giving up. Assertion count in both logs was `0 fail`. Re-runs cleared
+them, and the same commits pass the full suite on Linux.
+
+The `ci` aggregate job also failed on several branches by asserting before the
+long macOS job had reported; re-running it after macOS completes is the fix.
+
+## Screenshot gate
+
+`enforce-target` requires a UI screenshot for any PR touching `gui/`. These four
+change request lifecycle, timers, and caching with no visual delta, so the
+maintainer waiver label `gui-screenshot-waived` was applied to all four, and a
+post-change dashboard screenshot was captured during verification.
+
+## Merge
+
+Bottom-up, one at a time, each with `--match-head-commit` bound to the audited
+SHA. `dev` requires one approval and the author cannot self-approve, so the
+project owner's admin merge (pre-authorized by the user for this campaign) is
+the path used.
+
+- #1854 merged at 2026-08-16T17:15:18Z as `e2ef24ad6`.
+- Remaining layers merge in order once their own checks are green, retargeting
+  each child to `dev` after its parent lands.