fixed main thread hangs

[RaajeevChandran]

Summary

Twenty-one main-thread fixes from a Sentry app-hang triage. Every fix targets a verified main-thread stack (disk, keychain, database queue, or CPU work reached from a view body or the launch path), and each lands as its own commit so it can be bisected or reverted independently.

Launch

• The chat sessions manager's first metadata read is warmed through the database queue off the main actor before the speculative ChatView prewarm, so the prewarm no longer parks main behind a busy serial queue. Gated on storage readiness, matching the existing ChatView prewarm.
• Highlightr's JSContext (highlight.js) is warmed on a background thread during launch settle instead of booting on main when the first code block renders.
• The agent-secret account memo is seeded at launch so headless composers (HTTP, subagents, channels) never hit a cold keychain enumeration on main.

Chat and composer

• The voice button reads the speech service's published mic-permission mirror instead of a synchronous TCC round-trip on every body evaluation.
• The swap-emulation flag file is read on a background queue and memoized, so the 2s memory tick no longer touches disk on main.
• Chat model candidates are stored on the picker cache and recomputed per rebuild instead of string-filtered on every render.
• Token estimation uses UTF-8 byte length instead of grapheme counting, making the per-render context-budget math O(1) instead of O(conversation).
• Agent secret name listing on the compose path is non-blocking: a warm memo answers directly, a cold one kicks a single-flight background seed. The memo is maintained in place on save and delete so a store_secret mid-turn never causes the next send to see zero secrets.
• Live diff previews scan at most 256 KB of the streaming arg buffer per delta (byte-bounded), turning a quadratic total into a bounded one. Mid-stream syntax highlighting is capped by size with one plain rebuild at the crossing so later deltas don't inherit the last token's color.

Config apply and providers

• Keychain secret references resolve off the cooperative executor during declarative config apply. All resolutions are hoisted before the provider read-mutate-write so the suspension can never revert a concurrent Settings edit, and the remote-provider path re-fetches the live provider after the await.
• The relay WebSocket session (which reads the disk-backed proxy config) is built in a detached task. Connect re-checks by task identity across its suspensions and retires a stale socket safely, and the receive loop only tears down or feeds the current socket.

Downloads

• Download containment probes, resume size checks, the manifest completion check, and the compatibility report all run detached instead of on the main actor. The orchestration task inherits main-actor isolation from its service, so this code was on main despite reading like background work.

Quit

• The seven background-writer flushes in applicationWillTerminate drain concurrently under one bounded wait instead of serially on main. Keychain and config-disk drains get an explicit sub-cap so a slow securityd write still lands before the hard exit, and the group deadline stays under the app-hang watchdog.

Menu bar

• Status-item refreshes are coalesced to one pass per runloop turn instead of one per Combine publish.

Verification

• Builds green locally. Nothing here changes KV-cache prefix bytes: the prompt prefix hasher was deliberately left alone.
• Sentry: 30 issues marked resolved in next release with mechanism-naming reasons, 93 archived until-escalating after stack verification (zero first-party frames or swap-stall samples). Resolved issues auto-reopen as regressions if they fire after this ships.

Changes

• Behavior change
• UI change (screenshots below)
• Refactor / chore
• Tests
• Docs

Checklist

• I have read CONTRIBUTING.md
• I added/updated tests where reasonable
• I updated docs/README as needed
• I verified build on macOS with Xcode 16.4+