v0.3.0 — crawl timer and fetch retries

Added

• deploy/mcp-feeds-refresh.{service,timer} — the systemd units that crawl feeds, with the README section explaining why a deployment without them is broken. Nothing in the server crawls on its own, and a stale cache reports as an empty feed rather than as an error: the daily briefing ran for days on 61 unfetched feeds with no failure anywhere to notice. The timer fires hourly plus once at 05:45, because a consumer that windows on publication time permanently loses anything published between the last crawl and its own run.
• A fetch that hits a rate limit (429), a gateway 5xx, or a dropped connection is retried with exponential backoff, up to three attempts in all. A definite answer — 404, 401, 304 — is still acted on at once. Reddit throttles per exit IP for a few seconds at a time, which used to cost a feed its whole crawl.

Changed

• Documented HTTP_PROXY/HTTPS_PROXY/NO_PROXY, which httpx already honoured. They are the answer to a publisher that rate-limits your egress IP whatever User-Agent you send.