fix v1.44.2: external review CR-1..CR-3

CR-1: the lock invariant is restated precisely (never by an accidental
tap; the door card's labeled button is the ONE sanctioned surface),
unlocking now confirms, and smoke_lock_invariant exercises all five
actuation paths (icon tap, controls[], card entities, _cardToggle,
opening card).

CR-2: attachment migration is transactional — the server COPIES files,
the config is committed with its revision check, and only then the old
folder is removed via the new houseplan/files/cleanup. A rejected save
no longer leaves the stored urls pointing at an emptied folder.

CR-3: migrate returns an exact {source: written} mapping; only confirmed
copies are rewritten, destination name collisions get a unique name
instead of silently linking a pre-existing file, and a failed migration
raises a toast instead of being swallowed.

+1 unit test (119), +1 backend test, +1 smoke (51 total); docs
same-commit

docs/CHANGELOG.md

@@ -1,5 +1,28 @@
 # Changelog
 
+## v1.44.2 — 2026-07-27 (external code review: CR-1…CR-3)
+
+A second, adversarial review (of v1.44.0) produced three findings; all are
+addressed.
+
+- **The lock invariant is now precise and enforced (CR-1).** The reviewer was
+  right that "locks can never be actuated from the plan" was too absolute a
+  claim: the door card's Unlock button does call the service. That button is a
+  deliberate product decision, so the invariant is restated where it belongs
+  ("never by an accidental tap; exactly one labeled surface"), unlocking now
+  **asks for confirmation**, and a new smoke exercises all five actuation paths
+  to prove icons, `controls[]` and the device card still refuse locks outright.
+- **Attachment migration became transactional (CR-2).** Rebinding a marker used
+  to MOVE its files before the revision-checked config save — if that save was
+  rejected, the stored config kept the old urls while the files had already
+  left. Now the server **copies**, the config is committed, and only then the
+  old folder is removed (`houseplan/files/cleanup`).
+- **Failed or partial migrations no longer rewrite urls (CR-3).** The copy
+  reports an exact `{source: written}` mapping; only confirmed copies are
+  rewritten, name collisions get a unique name instead of silently linking a
+  pre-existing file, and a failed migration surfaces as a toast with the links
+  left pointing at the still-existing originals.
+
 ## v1.44.1 — 2026-07-27
 - Added the community chat everywhere users look: **https://t.me/ha_houseplan**
   (badge and header line in both READMEs, a "Getting help" section, the issue

docs/SCOPE.md

@@ -33,7 +33,7 @@ Editors are admin-only tools and must never leak interactions into View
 |---|---|---|
 | J1 | "Show the whole home and what's happening right now" — live spatial overview: device states, room fills (light/temp/LQI), values, multi-floor tabs | **Closed** |
 | J2 | "Something is wrong — show me *where*" — leak/smoke/gas pulse, open doors/windows, unlocked locks, red dot on devices HA added silently | **Closed** |
-| J3 | "Let me act on the obvious right from the plan" — tap-to-toggle for safe domains, info cards, guarded lock action (explicit button only, never a plan tap) | **Closed** |
+| J3 | "Let me act on the obvious right from the plan" — tap-to-toggle for safe domains, info cards, guarded lock action | **Closed** |
 | J4 | "From zero to a working plan in one evening, no Inkscape/YAML" — image/PDF/draw, floors-import wizard, room polygons bound to areas, curated auto-placement, editable icon rules | **Closed**; onboarding polish is *partial* (no registry-driven room suggestions) |
 | J5 | "Room climate at a glance" — per-room temperature/humidity, comfort-range fills, room-card metrics | **Closed** |
 | J6 | "Keep the plan true as the home evolves" — new-device flag, two editors, drag/resize, merge/split, multi-client live sync, optimistic locking | **Closed** |
@@ -52,6 +52,17 @@ Editors are admin-only tools and must never leak interactions into View
 ## Known gaps that fit the mission (build only on owner's request)
 
 - Person/presence shown in rooms (classic floorplan ask; pure J1).
+### The lock invariant, stated precisely (review CR-1)
+
+No lock or alarm panel is ever actuated **by a tap on the plan**: icons, lock
+badges, `marker.controls[]` and the device card all refuse (`resolveTapAction`
++ `TOGGLE_FORBIDDEN_DOMAINS`, `isControllable`, `_cardToggle`). There is exactly
+**one** sanctioned actuation surface: the labeled Unlock/Lock button inside an
+opened door card, which additionally confirms before unlocking. That is a
+product decision (2026-07-22), not an oversight — but it means the invariant is
+"never by accident", not "never at all". Any new actuation path must either
+refuse locks or be added to this paragraph.
+
 - Plan-level "security glance": one badge for "all locked / N open" (J2).
 - Threshold colouring for room-card metrics (J5).
 

docs/TESTING.md

@@ -40,6 +40,16 @@
       below; config/diagnostic entities are not listed; locks never toggle from
       the card [auto: smoke_card_controls]
 
+- [ ] Lock invariant, all paths (v1.44.2, review CR-1): icon tap, controls[],
+      device card and _cardToggle refuse locks/alarm panels entirely; the door
+      card's Unlock asks for confirmation, Lock does not [auto: smoke_lock_invariant]
+- [ ] Attachment migration is transactional (v1.44.2, review CR-2/CR-3):
+      rebinding COPIES files, saves the config, and only then deletes the old
+      folder; a rejected save leaves the old files and urls intact; a name
+      collision in the destination gets a unique name (the pre-existing file is
+      never silently linked); urls are rewritten only for confirmed copies
+      [auto: unit logic.test + tests_backend]
+
 ## Environments matrix
 
 Run the *core flows* (marked ★ below) in each environment at least once per minor release: