fix(desktop): reseed new chats from profile defaults, guarded against concurrent picks (supersedes #67207)

[OutThisLife]

Summary

Salvages #67207 (@deniqlab) for #65300 — a fresh Desktop chat should seed from the config.yaml default, not silently reuse the previous chat's sticky composer model. Adds the race guard @teknium1's review asked for.

Behavior

• New chat / fresh draft → reseed from profile default (refreshCurrentModel(true) on fresh-draft-ready + boot-onto-fresh-draft). This is the cross-surface-consistent behavior: CLI and Telegram already seed each session from config; Desktop's localStorage stickiness was the outlier (Hermes Desktop new session ignores config.yaml default and reuses sticky composer model #65300, and the earlier owl-alpha/[Bug]: Hermes Desktop ignores OpenAI Codex provider and routes GPT-5.5 requests through Nous Portal (CLI works correctly) #58498 thread).
• Incidental refreshes still never clobber a manual pick — the non-force path from fix(desktop): refresh default-derived composer model (supersedes #62486) #65896 is unchanged, so boot/focus/session-event refreshes preserve an in-context selection.
• A live session keeps its session-scoped model (guarded by $activeSessionId).

The race guard (teknium1's review)

@deniqlab's version used refreshCurrentModel(true), which bypasses both manual-state guards — so a picker selection made while the reseed's async getGlobalModelInfo() was in flight got overwritten by the resolved default, violating apps/desktop/AGENTS.md's "guard against the past: a stale response must never overwrite newer intent."

Fix: a monotonic manual-pick token (session.ts). selectModel bumps it; the reseed captures it before the await and bails if it moved while pending. So force still overrides a pre-existing manual pick (new chat resets to default), but a newer pick made mid-reseed wins. Added a deferred-result regression test proving the post-start pick survives.

Note on the #65896 relationship

This intentionally drops #65896's cross-chat manual stickiness (a pick no longer bleeds into the next new chat) while keeping its genuinely important part — the stale-guard against incidental refreshes. Net: manual intent wins within a chat and against in-flight reseeds; a brand-new chat honors your configured default.

Test plan

• vitest run --project ui — 189 files, 1530 pass / 1 skip (incl. new deferred-pick guard test + use-background-sync.test.tsx)
• npm run typecheck — clean
• eslint changed files — clean
• npm run build — success
• Manual: pick a non-default model, Cmd+N → composer shows the config default; pick a model immediately as a new chat opens → your pick sticks (not overwritten by the reseed).

Fixes #65300

[OutThisLife]

Closing — after weighing the UX, we're keeping the sticky-last-pick behavior (#65896) instead of resetting new chats to the profile default.

Switching between existing chats is already session-scoped and unaffected; the only thing this changed was the new-chat boundary, and clobbering a deliberate manual pick on Cmd+N is worse for the "I switched models on purpose" workflow than the reset is worth. The actual reported breakage — a removed/stale default (owl-alpha, #58498) lingering and overriding config — is already fixed on main by #65896, which re-syncs default/legacy-derived composer state from config.yaml while leaving explicit picks alone.

The race-guard was only needed because of the force-reseed, so it's moot without it. Thanks @deniqlab for the solid work regardless.