chore(upstream): adopt the third 2026-08-19 T3 Code batch

[rynfar]

Adopts 10 of 11 upstream commits from f2d5fc91..a850895f. git cherry reported every one absent from Pylon, and the deferred register was empty going in.

The skip is the notable decision

#6420 hides OpenCode's plan agent when legacy plan mode is off — a setting Pylon does not have. Its planModeEnabled flag arrived upstream in #5551, which this ledger records as skipped at F6: "Pylon keeps plan mode as a first-class composer affordance." The cherry-pick conflicted in SettingsPanels.tsx and modelSelection.ts precisely because it tried to reintroduce the rejected Settings row, so adopting it would have dragged in the setting F6 deliberately refused.

Skipped as a dependent of a skipped commit. The same reasoning should apply to future #6420-adjacent work.

The one substantial conflict

#7153 extracts the sidebar footer into SidebarUtilityMenu, generalizing onFooterPage into currentFooterPage. Pylon's footer had diverged 47/−24 through three fork-only commits — compact footer actions, the usage-page back button, and the Claude drain pill. I took upstream's side on all five hunks, after confirming its SidebarUtilityItem is markup-identical to Pylon's compact row and that SidebarAccountDrainPill sits outside the conflicted block.

That resolution then broke typecheck twice and no test caught it — nothing covers SidebarChrome.tsx. Taking upstream's import line dropped useId, which Pylon's own PylonMark cube logo uses, and canGoBack ended up declared twice. Fixed in d5eb43421. For a file where fork divergence is behavioral rather than textual, typecheck is the only gate.

Verification

• 17 changed test files, 238 tests, 0 failures.
• Typecheck clean across web, desktop, server, shared, and contracts.
• An xhigh review produced 14 findings. Three fixed; one investigated and rejected; the rest recorded in the ledger as upstream implementation choices.

Fixed from the review

• Usage listed every supported provider, used or not. #7147 switched the summary column from merged.providers to PROVIDER_ORDER. This lands worse in Pylon than upstream — we made Cursor, Grok, and OpenCode disabled by default yesterday, so a normal install now buries its one real provider under four 0 sessions · $0.00 rows.
• A second T3-branded line in docs/user/source-control.md that I missed on the first resolution pass.
• handleUsageClick still inlining the mobile-sidebar close that #7153 extracted.

Rejected, with reasoning

The reviewer called the install button vanishing during an update check a regression from #6269. Half right: #6269 does remove the guard that skipped checks while an update was downloaded, so the poller now runs one every four minutes and the button does blink. But hiding install while checking is older, explicitly tested behavior — desktopUpdate.logic.test.ts asserts "hides the install action while checking for a newer release" — and re-checking a queued update is the entire point of #6269. My attempted fix broke that test, which is the correct signal. Left as upstream's trade.

Six further findings are recorded in the ledger rather than fixed here, including an ElectronMenu separator bug that is latent today, and "Close others"/"Close all" skipping the terminal close confirmation #7592 just added.

Model: Claude Opus 5 in Pylon.

---

^{Need help on this PR? Tag @codesmith with what you need. Autofix is disabled.}