chore(upstream): adopt the 2026-08-20 T3 Code batch

[rynfar]

Adopts 6 of 8 upstream commits from a850895f..beab6886f. Two are deferred as DEF-5, with reasoning below.

#7602 closes the canary follow-up the second 2026-08-19 batch opened. @clerk/electron goes to stable 0.0.34, passkeys back to 0.0.3, and all four platform-specific native canary entries leave minimumReleaseAgeExclude. Verified against the exact condition the ledger recorded — npm view @clerk/electron dist-tags now reports latest: 0.0.34. Pylon stops shipping prerelease native binaries, so the "only darwin-arm64 was ever exercised" risk retires with it.

Also in: #7595 (usage hourly ordering), #6222 (terminal pane gutter), #7633 (test assertions), #7642 (Open VSX theme imports), #7151 (tool lifecycle identity).

A conflict resolution leaked a secret, and Pylon's own test caught it

Adopting #7151 as a plain union put upstream's unconditional toolCallId: event.itemId beside Pylon's primeAgentTool gating rather than behind it. A Prime tool's itemId is a canonical filesystem path — exactly the native detail that gating exists to withhold — so the projection began carrying canonical-prime-tool-/private/native-secret. ProviderRuntimeIngestion.test.ts asserts that string must not appear and failed immediately. Fixed in 0d70d9aec, along with two other gaps in the same resolution (upstream's status for non-Prime tools, and data on tool.started).

#7150 and #7152 are deferred, not skipped

#7150 rewrites 1084 lines of ChatComposer.tsx — 50 Pylon commits, 4837 lines against upstream's 2824. I made a full pass: classified all 11 conflict blocks, traced each one's provenance, and resolved them individually — keeping Pylon's settle loop (ce5371d41, which fixed a bug upstream's approach reintroduces), resolvedRuntimeMode, the 169 lines of Quick question and session-resource controls, and the ThreadHandoffTab wrapper, while taking upstream's drawer measurement and its 335-line drawer structure.

It compiled to 8 JSX errors. Upstream flattens the fragment and nested provider-frame divs into a single <form>, so its added tree cannot be hosted by Pylon's layers — the two are ends of one restructure. #7152 then conflicts on MessagesTimeline because it genuinely depends on #7150.

That needs the composer rebuilt on upstream's structure with Pylon's features grafted back, which is its own piece of work rather than a merge. Registered as DEF-5 with the full graft list.

Verification

• 115 tests across the 4 changed test files, 0 failures. Typecheck clean across 7 packages, each confirmed to have actually run.
• Live in the web client against seeded real data: pairing, thread view, the Usage page and its "Hourly cost by provider" chart, Settings → Appearance with the theme list and Import theme. No error surfaces.
• Not verified: #7595's chronological hour ordering. The chart exposes no hour labels to the DOM, so ordering could not be asserted from the browser — said plainly rather than implied.
• An xhigh review returned 13 findings. Two fixed in 2cf95d544: a status ternary whose arms were provably identical, and the missing test for a dedupe branch this batch widened.

Worth knowing, recorded in the ledger

#7151 widened the snapshot dedupe — toolLifecycleIdentity now prefers a payload-level toolCallId, which that commit sets on every tool activity. The clients read only data.toolCallId, so they can't mirror it: in-flight rows with differing details render live and collapse after a reload. The new test pins the behavior; whether it's the behavior Pylon wants is a separate call. Four more findings (unprojected tool.started data, a separator-less identity join, the relaxed zip-bomb ceiling in #7642, an unreachable Past-24h empty state) are in the ledger.

Model: Claude Opus 5 in Pylon.

---

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

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	12.6 KiB	13.4 KiB	+827 B (+6.4%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.3 KiB	6.9 KiB	+615 B (+9.6%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.3 KiB	6.5 KiB	+212 B (+3.3%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	51.8 KiB	55.0 KiB	+3.2 KiB (+6.1%)	66.4 KiB	✅
Codex	Live turn messages	16	16	0 (0.0%)	21	✅
Claude	Total thread wire	12.6 KiB	13.5 KiB	+877 B (+6.8%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.3 KiB	6.9 KiB	+627 B (+9.8%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.3 KiB	6.6 KiB	+250 B (+3.9%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	52.7 KiB	55.8 KiB	+3.2 KiB (+6.0%)	66.4 KiB	✅
Claude	Live turn messages	16	16	0 (0.0%)	21	✅

Baseline: f89e16d · PR result: 4a5f99a · Source CI: failure

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 109.4 KiB
• Claude decoded thread snapshot: 110.0 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.