Warm channel heads in the background before they are opened

[comp615]

What

Sessions created with warm: true (enabled in provideRelay, alongside prepared: true) warm every roster channel's head after discovery, so opening a channel — even after a cold start — finds its timeline already local:

• Ordering: starred channels first (from account sidebar preferences, now loaded at session start rather than waiting for the sidebar to mount), then the rest by recency of their retained head, never-fetched last. Restored disk heads are evaluated at drain time, so hydration finishing after discovery still reorders warmth by recency.
• Politeness: one background-priority read at a time (never competes with demand reads for foreground slots); skips channels with an open window (demand owns those); skips heads inside their 60s freshness lease — which also makes repeated roster refreshes a no-op until leases expire, answering the earlier requery-on-every-event concern.
• Avatar warming: each completed background head read also warms that channel's avatar URLs via media.warm(), which appends to the speculation queue after profile enrichment resolves instead of displacing focused-channel intents (capped queue, existing two-request concurrency).
• Save-Data: when the browser reports navigator.connection.saveData, all speculative warming is disabled — the hover speculative head read, the background drain, and avatar preparation. Demand fetches are untouched.
• Lifecycle: pending warmth is dropped when the session resets or clearCache() runs (so the Developer-tab clear-cache button from Add Developer settings tab with cache clear and broker stats #53 exercises a true cold load).

Sessions without the option keep the existing, test-encoded invariant that discovery never reads unopened channel heads — the existing relay tests pass unchanged; the new behavior is opt-in at the same boundary as prepared.

Validation

• New src/features/relay/warm.test.ts (deterministic scripted transport, no sleeps): cold-load auto-warm with serial ordering; starred/recency/never-fetched order; freshness-lease skip on re-trigger; open-channel skip.
• Media tests: prefetch warming appends across channels instead of replacing the queue; Save-Data disables all warming with zero requests.
• Updated the edge-pill browser journey whose baseline assertion ("only focused-row preparation, never a roster fanout") raced the now-intended background warm drain; it now asserts its durable intent directly — every read after the baseline targets a distinct channel, so cue and scroll interactions never re-read. Verified locally on chromium and webkit projects.
• Full relay suite: 503 passed (pnpm vitest run src/features/relay); pnpm typecheck and biome clean on touched files.
• Live-network verification on the real community relay: cold load produced 26 avatar thumb requests from background warming with no channel opened; opening warmed channels served 7 of 8 avatar loads from cache.

Deferred checks

• Live cold-load timing against the real relay (start → fully warmed) is best observed after deploy; unit tests cover ordering and skipping, not wall-clock warm duration.
• Very large rosters (500-channel discovery cap) warm serially; if that's too slow in practice, a bounded "warm the N most relevant first" slice is a small follow-up.

🤖 Description written by AI (Amp); commits signed off by the agent on Charlie Croom's behalf per DCO.

[baxen]

Nice nice