Audit HP-1540-02: the recorder chose the map id with an or-chain, so a
valid numeric map_index=0 fell through to selected_map (or 'default') and
the server stored the run under a key the renderer never looks up. The
choice is now resolve_map_id() — the explicit backend half of the contract
shared with vacMapIdFromAttrs in src/vacuum.ts: the first value that is
not None wins, zero and empty string included.
HP-1540-03: pairs was a plain source -> (marker, vacuum) dict, so the
second placement of the same robot (the documented two-floor case) evicted
the first and its server history silently stopped. A source now maps to a
list of pairs, every marker gets its own copy of the run, and the state
subscription set is deduplicated.
HP-1540-05: every config/set spawns async_refresh as a detached task; two
of them interleaving across the awaited config load could both subscribe,
overwriting one unsub handle — a callback leak until HA restart. Refresh
is serialized with an asyncio.Lock and teardown flags the recorder closed
first, so a refresh parked on its await can never resubscribe afterwards.
Regressions cover map_index 0/'0'/''/selected_map cross-checks, one
source feeding two floor markers across a map switch, pair-list refresh
with a deduplicated entity set, and two overlapping refreshes leaving
exactly one live subscription (zero after teardown). On the v1.54.0
recorder 11 of these tests fail.
At startup the vacuum entity reads unavailable; the recorder took that
for 'stopped' and ended the open run, so every HA restart mid-cleanup
rotated the trail into previous and began a fresh one (observed live:
a 21-point run became previous and restarted at 5). Unavailable and
unknown now mean 'no verdict'.
Caught live on the owner's X50 mid-cleanup with temporary logging: the
recorder saw every camera state change and rejected every one, because
server-side Tasshack keeps vacuum_position as a Point OBJECT — it only
becomes a dict when serialised to the frontend, which is why the card
adapter (and MCP inspection) always saw a dict and the stub test
faithfully reproduced the same wrong assumption. getattr fallback added,
regression test pinned, diagnostic logging removed (setup line kept at
INFO).
Only TrailBook was covered; the HA-facing half — subscription callback,
attribute dialects, map-id resolution, run end on docking — had no test
at all. It does now, against a stubbed hass, which is also where the
missing behaviour showed up: recording started at the NEXT state change,
so an HA restart (or finishing calibration) mid-cleanup dropped the
opening seconds of the path. Sampling is factored out and runs once per
source on setup and on every refresh.
The integration now records the path itself (trails.py): it watches the
source entity's state changes, so recording needs no open card, has no
multi-tab write races, and every screen sees the same line — reloads
included, which retires the localStorage snapshot after one day of
life. Stored per marker: the current run plus exactly one previous
(owner call — users want cleaned-vs-uncleaned at a glance). The
previous run renders at 40% opacity; the current one still trims its
live tail so it never outruns the puck. Runs rotate on start or map
switch, points cap at 2000 with decimation, store writes debounce 10 s,
and houseplan_trail_updated pushes live cards. TrailBook is pure under
5 backend tests; the WS command degrades silently on older backends.
Matysh