Owner's report: «план перезагружается при возврате на вкладку, хотя страница
жива». Diagnosis confirmed: Lovelace re-creates the card element when the
websocket reconnects after a long-backgrounded tab, and the fresh instance ran
the FULL first-open boot — veil + BOOT_MIN_MS + quiescence — reading as a plan
reload. On top of that, _loadFromServer's catch nulled _serverCfg after 8
failed tries, so a slow reconnect could genuinely blank an already-shown plan.
DEV-B703-01 warm re-mount: module-scoped memo (lives with the PAGE, not the
instance) of the settled header height, keyed by viewport size × card config.
A repeat instance adopts the settled geometry in setConfig and skips the veil
entirely — synchronous reveal at the saved zoom (HP-1551); a window resize
between instances changes the key and brings the full protective boot back.
The memo follows the live geometry (updated()'s measure) and is written on
every _bootSettled. Test hook: static _warmBootReset().
DEV-B703-02 stale-while-revalidate: an instance that already renders a valid
config (LS snapshot or a successful load) NEVER clears it on WS failures —
the local-only fallback is reserved for a card that never had a backend. A
self-driven retry (backoff, cap 8 s) keeps revalidating after willUpdate's
8-try budget is spent, and a connection 'ready' hook resets the budget and
quietly re-reads the config the moment the socket is back (the event
subscriptions re-subscribe on their own inside home-assistant-js-websocket).
Smokes: smoke_warm_remount (fails pre-fix: veil + hidden plan on re-mount;
resize invalidation stays cold), smoke_ws_resilience (fails pre-fix:
_serverCfg cleared after 8 tries, no revalidation after recovery); the
preloader smokes now reset the warm memo — they simulate a COLD first open.
The v1.55.2 recheck (verdict NEEDS FIX) found two lifecycle holes in the
first-open boot veil (HP-1552); both are closed here and covered by
demo/smoke_preloader_lifecycle.mjs, which fails on v1.55.2.
AUD-1552-01 (high): disconnect/reconnect while booting hid the plan
FOREVER. disconnectedCallback cleared the boot timer but kept its
truthy id, so 'updated()' never restarted the watcher. Now the id is
nulled on disconnect and connectedCallback restarts the whole veil
lifecycle: a fresh watch (fresh clock, BOOT_MAX_MS hard cap) while
booting, the tail timers when detached mid-fade or mid-grace.
AUD-1552-02 (medium): two equal reads at 200/400 ms revealed the plan
at ~400 ms, so HA chrome landing at 450+ ms jumped on a VISIBLE plan.
The veil now holds a full protective window (BOOT_MIN_MS=700 — the old
600 ms plus a frame-latency margin: a shift applied at ~590 ms only
materializes in the stage height a couple frames later) with 100 ms
sampling and trailing quiescence (BOOT_QUIET_MS=250 restarts on every
height change), capped by BOOT_MAX_MS=1200. After the reveal a short
soft grace (BOOT_SOFT_MS=1500, .stage.hpsettle) turns later passive
shifts into a 0.25 s height glide — the viewport ResizeObserver refits
the plan along the transition; deliberate height changes (_setMode into
an editor) cancel the grace so the plan never drifts under the pointer.
Reduced motion disables the glide.
Regressions (demo/smoke_preloader_lifecycle.mjs, all FAIL on v1.55.2):
- A: detach before the first tick -> reattach -> veil lifts within the
cap, plan visible, no hpboot class, no zombie veil after a mid-fade
remount either;
- B: parameterized layout shifts at 150/300/450/590 ms -> not a single
frame shows the plan at a non-final stage height;
- C: a shift after the reveal glides (>=3 intermediate frames), no snap.
smoke_modes.mjs now strips the transient hpsettle class from its exact
stage-class assertions. Docs: CHANGELOG en+ru, STATUS.
Matysh