upstream: adopt the clean half of the 2026-09-03 T3 batch (65 commits)

[rynfar]

Upstream moved 151 commits since Pylon's cursor at beae2147a. The maintainer
chose to adopt everything except T3 marketing content, so this is pass one of
two: every commit that applies cleanly, landed in upstream order.

What is here

65 upstream commits, plus 4 Pylon-side commits that reconnect the fork.

The interesting part is that "applies cleanly" is not the same as "is correct
for Pylon". Three upstream changes produced no conflict markers and still broke
the build, because each was written against a tree that does not contain
Pylon's files:

• #7272 moved the node:sqlite Effect SQL client into shared and updated its
  own importers, leaving Pylon's migrations 043/044/047/048/049 and the Prime
  Agent recovery ledger test pointing at a deleted module.
• #9210 landed and returns snapshotForCwd, but the commit that declares it
  on ProviderInstance (#8778) is still parked.
• #8909's new provider-session directory mock predates Pylon's removeExact,
  which Prime Agent admission recovery added.
• #9258 calls isAbsolutePath, which upstream exports from mobile's
  filePath.ts and Pylon kept module-private.

#9129 ("delete dead code") needed the most care. It computed liveness against
upstream's tree, and several helpers it deletes are still called here —
resolveProviderHomePath has 13 production callers, four of them Prime Agent;
getModelSelectionOptionValue resolves Prime Agent thinking level and service
tier; mergeProviderCapacityRefresh is used inside ProviderRegistry itself;
and the desktop preview recording schemas are live across the preview stack.
The deletions that are genuinely dead here were taken; the rest are restored in
a separate commit so the carve-out is reviewable on its own.

Two commits that had already landed clean, #9306 and #9279, are reverted
here: both depend on parents still parked (#9146 citations, #7254 browser
profiles), and a branch that does not compile is not worth carrying.

Not here

• #9385 (marketing landing-page quotes) is skipped — every quote names T3 Code.
• #8808 + #9284 are skipped as a pair. Upstream added context compaction and
  reverted it the same day, so adopting both yields a byte-identical tree.
• 86 commits conflict and are parked for pass two, including the four migration
  renumbers, #9348 Antigravity, and the vite-plus 0.3.0 bump.

Verification

• typecheck clean across web, mobile, desktop, server, contracts,
  client-runtime and shared.
• Focused tests green: 159 passed across the touched server and client-runtime
  suites, 24 across the touched mobile suites. One run of the server batch
  showed two failures in ProviderRegistry.test.ts that did not reproduce on
  re-run or in isolation (57/57); recorded as a flake, not a fix.
• Note that both vp run and vp test run exit 0 while reporting failures, so
  these were read from output rather than exit status.

The ledger cursor is deliberately not advanced — that happens once pass two
resolves the parked commits and every candidate has a decision.

Made with Opus 5 in Claude Code.

---

^{Need help on this PR? Tag @codesmith-bot 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	13.6 KiB	13.6 KiB	+13 B (+0.1%)	15.1 KiB	✅
Codex	Thread snapshot wire	7.1 KiB	7.1 KiB	+1 B (+0.0%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.5 KiB	6.5 KiB	+12 B (+0.2%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	56.5 KiB	56.5 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	8	8	0 (0.0%)	21	✅
Claude	Total thread wire	13.6 KiB	13.6 KiB	+20 B (+0.1%)	15.1 KiB	✅
Claude	Thread snapshot wire	7.1 KiB	7.1 KiB	+1 B (+0.0%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.5 KiB	6.6 KiB	+19 B (+0.3%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	57.3 KiB	57.3 KiB	0 B (0.0%)	66.4 KiB	✅
Claude	Live turn messages	8	8	0 (0.0%)	21	✅

Baseline: b8e5774 · PR result: 95a169e · Source CI: success

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: 111.2 KiB
• Claude decoded thread snapshot: 111.9 KiB

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