perf(channels): stale-while-revalidate for region-filtered channel lists (cold 15 s after #499) #516

Problem

After #499 (region-correct channel counts), the cold region-filtered channel list got slower, as #499 disclosed (+34–57%). Measured on staging v0.2.6-rc1 against production v0.2.5:

The channel-list cache has a 60 s TTL with singleflight. The first request after expiry waits for the full cold query, so roughly one user per minute per region can wait up to 15 s.

Task

Add stale-while-revalidate to the channel-list cache, covering GetChannels(region) and GetEncryptedChannels(region), as proposed in #379:

• When an entry has expired, serve it immediately and start one bounded background refresh per key, coalesced with singleflight.
• Only a truly empty cache (first request after start, or an evicted key) waits synchronously.
• Keep the cache keys bounded, keep error handling (a failed refresh keeps the stale entry and retries later, without a stampede) and keep the server read-only. No SQL under PacketStore.mu (perf(store): move runtime SQL outside packet-store locks #495 rule).
• Bound staleness. For example, serve stale for at most N× TTL, and wait synchronously after that.
• Optional: speed up the cold query itself (query plan or index), if that can be done without breaking the fix(channels): keep regional counts and latest message consistent #499 correctness guarantees.

Tests and evidence

• Test that an expired entry is served immediately while exactly one refresh runs under concurrent requests.
• Test the bounded-staleness cutoff.
• Test that a failed refresh keeps serving stale.
• Measure before/after locally on a realistic fixture. The coordinator re-measures on staging.

Acceptance criteria

After warm-up, no region-filtered channel-list request waits for a cold query on a cache expiry. The #499 correctness tests stay green.