docs: a changelog, and the release checklist says to write in it

Every release so far has had its notes on the GitHub releases page and
nowhere in the tree, so somebody standing in a checkout could not find out
what they were about to take without leaving it. This adds CHANGELOG.md,
Keep a Changelog format, starting at 0.6.0; earlier versions keep pointing
at the releases page rather than being reconstructed from commits.

The entry opens with what an upgrading operator has to do by hand — for
0.6.0 that is nothing — because that is the line people look for, and a
changelog that buries it is a list of commits with better formatting.

RELEASING.md gains the step, and says the GitHub release notes are this
entry pasted: two places disagreeing about one release is exactly the drift
that file exists to prevent.

Author: bridgerahul

CHANGELOG.md

@@ -0,0 +1,79 @@
+# Changelog
+
+The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
+and the numbers mean what [RELEASING.md](RELEASING.md) says they mean: while
+the major is `0`, a minor is anything you might have to read this file before
+taking.
+
+Releases before 0.6.0 have their notes on the
+[GitHub releases page](https://github.com/OpenLabs-so/openanalytics/releases).
+
+## [0.6.0] - 2026-09-05
+
+**Upgrade notes: none.** `./upgrade.sh` and nothing else. The two ClickHouse
+migrations in this release only add columns, and the migrate container applies
+them while the upgrade runs.
+
+### Added
+
+- **Filters.** Clicking a row in Sources, Locations or the device cards keeps
+  only the sessions that match it, and the overview re-reads with that filter
+  applied. Four dimensions — source, country, city, device — combined with
+  `and` across dimensions and `or` within one. What you picked shows as marks
+  on the tab bar, each clearing on its own. A filtered view is rebuilt from raw
+  events rather than read off a rollup, so it covers at most 92 days; past that
+  the cards offer to shorten the range or drop the filters. The same grammar
+  serves the dashboard, `/v1/read` and MCP, so they cannot drift.
+- **Entries and exits.** Top Pages carries entrances, exits and a bounce rate
+  per path, and can be ranked by any of them. The ranking is a question put to
+  the server rather than a re-sort of the rows on screen.
+- **Click ids as sources.** `gclid`, `fbclid`, `msclkid` and their companions
+  name the ad network a visit came from when the referrer does not.
+- **`?ref=` names a source.** A link you control can carry `?ref=twitter` or
+  `?ref=newsletter`, and visits that would have read as Direct get that name
+  instead. Known names fold onto the site they mean; anything else is kept as
+  written. A referrer the browser did send always wins.
+- **Polar as a revenue provider.** A second provider a site can connect
+  alongside Stripe, with Standard Webhooks signature verification.
+- **Umami import.** An adapter for Umami exports, and one archive entry may now
+  fill several reports.
+- **Overview poster.** A share button turns the numbers on screen into an
+  image — the site's mark, the period, the headline figures and the traffic
+  curve — as a PNG or on the clipboard. It draws from what the board already
+  has, so it cannot disagree with the screen it was made from.
+- **Realtime names who is here.** The overview's realtime card lists the
+  visitors present rather than the paths that are busy, and a live badge sits
+  beside the Overview and Realtime titles.
+- **MCP hints.** Every tool spells out its hints, native-app redirect URIs are
+  accepted, and a root `.mcp.json` points at the hosted server.
+- **Dokploy.** The blueprint ships with an import config and a copyable raw
+  link beside it.
+- Documentation for the dashboard: what each headline metric counts, what each
+  breakdown is a cut of, how filters behave, and how to tag your own links.
+
+### Changed
+
+- **The tracker stands down for a site that is gone.** A `404` from the config
+  read is a gone-marker, and `collection_paused` carries a body and an ETag;
+  the tracker obeys both and stops asking.
+- Squircle cards fall back to rounder corners on engines without
+  `corner-shape`, and to a plain radius where even that is unsupported.
+
+### Fixed
+
+- **Email.** A rate limit is retryable and a bad address is not, so a `429` no
+  longer burns an address; a stranger's typo no longer becomes a page. A
+  terminal failure is recorded as `failed`.
+- **Ingest.** The finalizer window is bounded, the outbox is leased and
+  reclaimed rather than raced, the backlog carries an `available_at`, and the
+  queue publishes its own metrics.
+- **Realtime.** A refusal is believed only once a fresh token has been refused
+  too, so an expired token no longer reads as a revoked one.
+- **Shared reports.** Percentages are shares of the total rather than of the
+  biggest row.
+
+### Coming in 0.7.0
+
+Fifteen-minute rollups, which serve every timezone including the half- and
+quarter-hour ones (+05:30, +05:45), with a one-time backfill the upgrader runs
+for you.

RELEASING.md

@@ -113,8 +113,13 @@ keepable. Not before.
 5. Watch it. **A release whose workflow failed is a tag pointing at images that
    do not exist**, and the failure mode for whoever tries it is a `docker
 compose pull` that cannot find a manifest.
-6. Write the release notes on the GitHub release. What changed, and — first,
-   before anything else — whatever an upgrading operator has to do by hand.
+6. Add the entry to `CHANGELOG.md` in the same commit as the version bump. It
+   opens with what an upgrading operator has to do by hand, or with
+   **Upgrade notes: none** — that line is the first thing somebody looks for
+   and the reason the file exists rather than a list of commits.
+7. Write the release notes on the GitHub release. They are the CHANGELOG entry
+   for this version, pasted: two places that say different things about one
+   release is the drift this file exists to prevent.
 
 **Pre-releases** are `vX.Y.Z-rc.N`. They publish the same ten images under the
 pre-release tag and deliberately do not move `latest`, so the way to test the