docs(release): notes and changelog for v3.13.0 (#2110)

Release notes and CHANGELOG entry for v3.13.0, in the shape used for
v3.12.0. Once merged, the tag goes on the merge commit so `deploy.yml`
picks up `docs/release-notes/v3.13.0.md` as the release body (#2076).

- 16 commits since v3.12.0: 10 fix, 4 test, 2 feat, so a minor bump.
- "Read this before upgrading" covers the one-off advert route evidence
backfill from #2088. Durations are measured on two instances (live 24m
1s on 16.8M observations, staging 36m 23s on 16.5M), and ingest during
the backfill window is measured as unaffected on live.
- Issues closed come from each PR's closing references: 14 from 16 PRs.

Docs only. #2089 is deliberately not in this release.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

CHANGELOG.md

@@ -2,6 +2,19 @@
 
 ## [Unreleased]
 
+## [3.13.0] - 2026-10-04
+
+See [docs/release-notes/v3.13.0.md](docs/release-notes/v3.13.0.md) for the full notes. 16 commits since v3.12.0: 10 fix, 4 test, 2 feat.
+
+### Highlights
+- **The first start backfills advert route evidence in the background, and flood counts read low until it finishes** (#2088) - measured at 24m 1s on 16.8M observations and 36m 23s on 16.5M on a busier host. Startup and ingest do not wait for it. Runs once per database and resumes after a restart. **Operator awareness required.**
+- **Adverts are split by the route they arrived on** (#2088, #2085) - flood, direct with an empty path, or mixed, in Analytics and on node detail.
+- **Path hops are resolved with the observer that heard them** (#2099) - with one per-path warning in the packets list when hops stay ambiguous.
+- **A compact packets table with a Full Names toggle** (#2090).
+- **Channel unread counts and key-only channels survive a list refresh** (#2096).
+
+No manual migration step: `advert_route_evidence` is created at boot.
+
 ## [3.12.0] - 2026-09-26
 
 See [docs/release-notes/v3.12.0.md](docs/release-notes/v3.12.0.md) for the full notes. 29 commits since v3.11.0: 15 fix, 5 test, 3 perf, 3 feat, 2 ci, 1 chore.

docs/release-notes/v3.13.0.md

@@ -0,0 +1,131 @@
+# v3.13.0
+
+16 commits since v3.12.0: 10 fix, 4 test, 2 feat. Two feats, so minor not patch.
+
+## Read this before upgrading
+
+**The first start after this release backfills advert route evidence in the
+background, and flood counts read low until it finishes.**
+
+#2088 records, per advert, whether it was seen arriving by flood, by direct with
+an empty path, or both. New adverts get this at ingest. For the adverts already
+in your database, the ingestor runs a one-off background job
+(`advert_route_evidence_v1`) that reads their stored observations. Until it
+completes, the Analytics advert split and each node's flood-advert count only
+know the adverts it has reached so far, so they undercount.
+
+Measured on two instances of the same deployment, both on a 4-core ARM64 host:
+
+| | observations | database | duration |
+|---|---|---|---|
+| live | 16.8M | 10.6 GB | 24m 1s |
+| staging, sharing the host with live | 16.5M | 10.3 GB | 36m 23s |
+
+The author measured 5m43s on 11M observations on other hardware, so expect a
+wide spread.
+
+Startup does not wait for it, and ingest carries on alongside it: on the live
+instance the 15 minutes from 13:00, inside the backfill window, stored 2,803
+observations against 2,605 to 3,194 in the quarter hours before the deploy.
+
+It runs once per database. Its progress is kept in the database
+(`advert_evidence_backfill` holds the cursors, `_async_migrations` the status),
+so a restart resumes rather than repeats it. There is no setting to skip it.
+
+What it cannot recover: an observation that was overwritten before this release
+keeps only its last frame, so some old adverts stay unclassified. The Analytics
+page says so under the chart.
+
+## Features
+
+- **Adverts are split by the route they arrived on** (#2088, closes #2041) -
+  Analytics separates flood adverts from direct adverts with an empty path, and
+  shows adverts seen both ways as mixed. Node flood-advert counts use the same
+  evidence. An empty remaining path is labelled as exactly that: the firmware
+  can forward a direct advert down to an empty path, so it does not prove the
+  advert was sent zero-hop or from nearby. If writing the evidence fails, the
+  error is logged and counted and the observation is still stored.
+- **Node detail groups recent adverts the same way** (#2085, closes #2073) -
+  both the side pane and the full page group by the server's classification.
+  An advert without evidence goes to "other" rather than being guessed from the
+  first stored frame, which could belong to a different route than later
+  observations of the same advert.
+
+## Paths and packets
+
+- **Path hops are resolved with the observer that heard them** (#2099, closes
+  #2097) - a 1-byte hop prefix matches several repeaters. Resolution now uses the
+  receiving observer's position, and the packets list shows one warning at the
+  start of a path saying how many of its hops have more than one candidate. The
+  detail pane keeps a badge on each such hop, tied to its name.
+- **The packets table is compact** (#2090) - short columns such as time, size
+  and hop count are sized to their content instead of a share of the screen, a
+  grouped row draws one expand arrow, and a **Full Names** toggle shows observer
+  and hop names unabbreviated.
+- **Packet links keep the selected observation** (#2093, closes #2091) -
+  changing the type or observer filter no longer drops `?obs=` from the URL.
+
+## Interface
+
+- **Channel state survives a channel-list refresh** (#2096, closes #2095) -
+  unread counts, last-message previews and the channels you added with a key
+  are kept when the region filter or the show-encrypted toggle reloads the list.
+  Before, every unread count reset and an open key-only channel was closed.
+- **Live accepts node-filter input only once it can process it** (#2103, closes
+  #2094) - typing while the initial node list loaded was silently lost.
+- **Heard By drops a Region column that was always empty** (#2077, closes
+  #2062) - the endpoint never sent that field.
+- **Navigation and channel buttons are one size, 48px** (#2078, closes #2052).
+- **The route sidebar no longer covers Path Inspector candidates** (#2082,
+  closes #2060 and #2081).
+
+## Tests, CI and release tooling
+
+- **The GitHub release body is these notes** (#2076) - v3.10.1, v3.11.0 and
+  v3.12.0 were published with an empty body.
+- **Four intermittent E2E and server-test failures fixed** (#2084, #2087, #2105,
+  #2109) - #2109 is the Path Inspector failure that turned master red twice on
+  2026-10-03.
+- **Packets fixture tests are stable, mobile empty states kept** (#2080, closes
+  #2054).
+
+## Issues closed
+
+14 issues, from the 16 pull requests in this range. Collected from each pull
+request's closing references rather than from commit text.
+
+- #2041 [feat] #/analytics split advert into flood-advert and zero-hop-advert
+- #2052 style.css declares .nav-btn and .ch-icon-btn at both 48px and 44px
+- #2054 26 wired E2E suites depend on the packets page's 15-minute window surviving the length of the job
+- #2060 The E2E fixture's neighbour graph yields no path-inspector candidates, so two assertions cannot run
+- #2062 Node detail "Heard By": the Regions line and Region column read a field the endpoint does not emit
+- #2073 [feat] node detail, separate flood-adv from zero-hop adv
+- #2081 bug(map): route sidebar shifts Leaflet over Path Inspector candidates
+- #2083 test(server): node-path regressions race asynchronous index readiness
+- #2086 test(e2e): channel WS assertions race initial channel loading
+- #2091 Packets: changing the type or observer filter drops ?obs= from a packet detail URL
+- #2094 bug(live): node filter accepts input before its handler is ready
+- #2095 channels: loadChannels() replaces the channel array and discards all client-only state
+- #2097 packets: an ambiguous path hop is shown as a single certain name, resolved without the geographic filter
+- #2104 test(packets): observation hex test clicks stale DOM handles
+
+## Contributors
+
+Pull request authors: @n30nex (11), @efiten (4), @sylr (1).
+
+No co-authors other than tooling are credited in the commits.
+
+Issues in this release were reported by @efiten (6), @n30nex (5),
+@damn-simple-scripts (2), @dborup (1).
+
+@n30nex wrote 11 of the 16 pull requests, including both features. @efiten, the
+interim maintainer, wrote 4 and reviewed and merged the rest.
+
+## Upgrade notes
+
+No manual migration step. #2088 adds the `advert_route_evidence` table, created
+at boot, and the background backfill described at the top fills it for existing
+adverts. It needs no action, only the awareness that flood counts are low until
+it finishes.
+
+Nothing is opt-in and no configuration key was added or changed.