chore: sync nightly v0.0.39-nightly.20260906.1292 into orchestration-v2

[charlieadair]

Brings 260 upstream commits (nightly .1279 → .1292) onto orchestration-v2 and reconciles them with the v2 orchestration stack. 108 files conflicted.

How conflicts were resolved

The branch's existing split held: upstream owns provider, UI and docs behaviour; this branch owns orchestration. Where both sides changed the same thing, these were the judgement calls worth knowing about.

Kept this branch's side

• Deleted v1 modules (adapters, reactors, ProviderService, ProviderSessionDirectory) stay deleted; upstream's edits to them are dropped.
• ws.ts, serverRuntimeStartup.ts, session-logic.ts, threads.ts and the mobile feed stay on their v2 shapes (runId/latestRun/runtime rather than turnId/latestTurn/session).
• Cursor stays on this branch's SDK path.
• The synthetic "Thinking" row upstream adds to the mobile feed stays out — this branch has a test asserting it is absent.

Adopted upstream's side

• The provider-driver contract change: resolveMaintenance replaces maintenanceCapabilities across Cursor, Grok and AcpRegistry.
• The persistent changed-files tree (fix(web): simplify changed files into a persistent folder tree pingdotgg/t3code#9821), replacing this branch's expand/compact-preview variant. Its call site and two stale assertions were updated to match.
• The thread-retention change (perf(client): stop thread streams when unused pingdotgg/t3code#9740): THREAD_SNAPSHOT_IDLE_TTL_MS moves off the raw stream atom onto the derived detail atoms.
• Docs: upstream's reorganisation (docs: keep internal guides focused on architecture pingdotgg/t3code#9755, docs: focus user guides on features and workflows pingdotgg/t3code#9756), with this branch's thread-migration note and composer-context section carried across and four provider links repointed at the v2 modules.

Ported rather than taken or dropped

• Warm thread resume (fix(client): keep warm thread resumes live instead of flashing sync pingdotgg/t3code#9913) onto this branch's history-based snapshot model. matchesThreadSnapshot now compares history.historyCursor/hasMoreHistory instead of the page cursor, and remember runs on the applied-state path so the persist guard compares against live state rather than a stale commit.
• Agent-session import onto v2. Upstream drives it through the v1 engine and session directory, neither of which exists here, and v2 has no history-import command. The thread is now created through ThreadManagementService and the transcript is written as ordinary message.updated/turn-item.updated events, mirroring LegacyV1ThreadImporter so no new persisted event type is introduced. EventSinkV2 and ProviderSessionRuntimeRepository moved into the production layer so the requirement does not leak to every CLI caller.
• historyImport metadata carried into this branch's consolidated ApplicationEventMetadata.
• The composer's new prompt-history navigation kept alongside this branch's Cmd+Enter queue gesture, which needs the modifier free.

Dropped

• Upstream test suites targeting APIs this branch removed: the v1 threads-atoms suite (restored to this branch's stub, updated for the new retention model), deriveTimelineEntriesWithState, deriveWorkLogEntries, upstream's older-page pagination tests, and one test driving an applyLock semaphore this branch does not have.
• buildRevertTurnCountByUserMessageId and selectHandoffImageResources, which had no production consumer here after resolution.

Verification

• vp run -r typecheck — clean, all 15 projects.
• client-runtime — 1306/1306.
• web — 3992/3992.
• apps/server/src/project/AgentSessionImporter.test.ts — new focused v2 test, 2/2.
• Touched mobile suites — 41/41.

Pre-existing and unrelated to this merge:

• Three apps/web/src/terminal/ghostty/* files fail to load on .wasm?inline. Two are byte-identical to the merge base, and the failure reproduces with the pre-merge vite configs.
• pnpm install fails in this environment with Cannot destructure property 'verified'. Root cause is WATCH_REPORT_DEPENDENCIES=1, inherited from the --watch dev-runner that hosts an agent session: every Node worker then posts a watch:require frame first, and pnpm's once("message") consumes it instead of the real response. Unsetting that one variable fixes it. Worth knowing for anyone running installs from inside T3 Code.

Follow-ups, not in this PR

• Re-port Cursor to main's ACP path. The SDK swap (8e2e8edd3, authored 2026-06-30) never landed on main, and main has kept improving the ACP path since. Retiring the SDK path means ~3,200 lines and re-recording ~20 testkit fixtures against ACP.
• The server suite had not finished when this was opened; anything it surfaces will be pushed to this branch.

Model: Claude Opus 5. Harness: Claude Code.