diff --git a/docs/specs/373-space-card-house-fit.md b/docs/specs/373-space-card-house-fit.md
new file mode 100644
index 00000000..50bd980a
--- /dev/null
+++ b/docs/specs/373-space-card-house-fit.md
@@ -0,0 +1,336 @@
+# #373 — Tight house framing for `houseplan-space-card`
+
+Issue: [#373](https://github.com/Matysh/houseplan-card/issues/373)
+
+## Scenario
+
+The **home administrator** embeds `custom:houseplan-space-card` in a compact
+Home Assistant dashboard for household members, a wall tablet or a phone. The
+stored space also contains a large backdrop or auxiliary objects outside the
+building, but this card is meant to show the house itself as large as possible.
+
+The administrator selects a tighter frame for this one card instance. View and
+kiosk consumers then see the complete structural plan without the ordinary
+outer breathing room; no editor is involved.
+
+## What the person sees before and after
+
+Before, the static card keeps a 5% outer frame and lets non-house content widen
+the view; after opting in, the complete house geometry touches the available
+frame without intentional padding, while backdrop, decor, labels and devices
+outside the house no longer make the house smaller.
+
+## Problem
+
+`renderSpaceStatic()` currently calls `spaceFrame(space, placed, 0.05)` (or the
+per-edge #372 variant for `title: ""`). That frame is deliberately the broad
+content frame from `docs/CANVAS.md`: rooms, the transformed backdrop and visible
+placed objects can all vote, and the result is padded by 5% of its longer side.
+This is correct for the existing default and for the full interactive card, but
+it wastes scarce dashboard area when the embedded card is intended to show the
+building rather than every auxiliary object.
+
+The reporter chose the non-cropping alternative in the latest issue comment:
+remove the 5% padding, keep all house geometry visible, and derive the fit from
+room/house geometry instead of backdrop or decorative objects. A CSS-`cover`
+mode that can crop the house is explicitly not requested.
+
+## Scope
+
+- only `custom:houseplan-space-card`;
+- one opt-in Lovelace card setting, `fit: house`;
+- visual-editor choice between the existing frame and the tight house frame;
+- a zero-intentional-padding structural frame containing every rendered room,
+  room wall, independent wall/draft, column, zero-thickness wall and visible
+  door/window/gate symbol;
+- exclusion of backdrop image, decor, room labels, device markers and their
+  badges from the `house` frame vote;
+- safe fallback to the existing frame when the space has no valid structural
+  geometry;
+- consistent geometry, themes and touch/desktop View behaviour;
+- en/ru/de/fr UI strings, RU/EN user guide, architecture/canvas contract,
+  changelogs and targeted tests.
+
+## Non-scope
+
+- any change to the full `custom:houseplan-card`, its Fit all action, View,
+  kiosk, editors, zoom, pan or outlier hint;
+- CSS `cover`, aspect-ratio distortion or any mode allowed to crop structural
+  house geometry;
+- hiding, deleting or moving backdrop, decor, labels, markers or devices;
+- changing marker size, room-label position, Glow, sunlight, wall geometry,
+  opening state or HA actions;
+- changing the stored House Plan model, backend schema, `space.view_box`,
+  config/layout revisions, exports or imports;
+- arbitrary user-configurable padding values;
+- changing the #372 meaning of `title: ""` in the default frame;
+- automatically enabling the new mode for existing cards.
+
+## Behaviour contract
+
+### Public card configuration
+
+```yaml
+type: custom:houseplan-space-card
+space: ground_floor
+fit: house
+```
+
+1. `fit` accepts `content | house`.
+2. Missing, empty or unknown `fit` resolves fail-safe to `content`.
+3. `content` is the exact current behaviour: the current content voters,
+   main-mass/outlier policy and 5% padding remain unchanged. Existing YAML
+   without `fit` must produce the same frame.
+4. `house` selects the new tight structural frame described below. It never
+   becomes the implicit default.
+
+### Tight structural frame
+
+5. The structural frame contains the union of all finite, sane visible
+   architectural extents for the selected space:
+   - room polygon/rectangle floors;
+   - the complete outer envelope of positive-thickness room walls;
+   - independent walls, saved wall drafts and columns;
+   - zero-thickness room and independent wall axes;
+   - the full visible state-independent envelope of door, window and gate
+     symbols, including wall jambs, leaf tips and swing arcs in either state.
+     An open passage contributes through its structural wall/tunnel geometry
+     and has no extra standalone symbol.
+6. All structural items participate. The broad content-frame outlier vote may
+   not discard a detached room or wing in `house` mode: sane structural
+   geometry is part of the house and must stay visible.
+7. The resolved structural union receives **zero intentional padding** on all
+   four sides. Bounds still include the actual rendered wall/symbol stroke
+   envelope, so “zero padding” may not clip a stroke by half its width.
+8. Backdrop image, decor, room labels, device markers, value/LQI badges, pulse
+   shells and purely visual Glow/sun extents do not enlarge this frame. They
+   remain rendered in the same coordinate system and may be clipped when they
+   lie outside the structural house frame; this is the user-confirmed trade-off.
+9. The SVG, marker layer, room-label layer, backdrop, day-cycle environment and
+   continuity overlay all consume the single resolved `viewBox`. Nothing is
+   independently translated or scaled after the fit is chosen.
+10. If no valid structural item exists (for example an image-only or empty
+    space), `house` falls back atomically to the ordinary `content` frame. It
+    must not create an empty, zero-sized, NaN or infinite `viewBox`.
+11. Structural computation is state-independent. Opening sensor changes,
+    device state updates, Glow animation and language/theme changes do not
+    resize the card or move the house inside it.
+
+### Interaction with existing options
+
+12. `title`, `show_button`, `light_pools`, state/value options and footer deep
+    links retain their existing semantics.
+13. With `fit: content`, explicit `title: ""` keeps #372: only the top 5%
+    padding disappears, while side/bottom padding remains.
+14. With `fit: house`, all four intentional frame paddings are already zero;
+    `title: ""` only removes the header and does not change the structural
+    `viewBox` further.
+
+## UX and accessibility
+
+- The Lovelace visual editor exposes one dropdown labelled **Framing** with:
+  - **All visible content** → `content`;
+  - **Tight to house geometry** → `house`.
+- A missing legacy value is displayed as **All visible content**. A subsequent
+  visual-editor save may materialise `fit: content`; this is semantically
+  lossless.
+- No crop/cover choice is shown.
+- The schematic remains `pointer-events:none`, non-focusable and read-only.
+  The existing footer button remains the only interactive element.
+- No new hover, gesture or animation is introduced. The frame is identical on
+  fine/coarse pointers, narrow/wide cards and light/dark themes.
+
+## Data model, migration and compatibility
+
+The field belongs only to the Lovelace card instance:
+
+```ts
+interface SpaceCardConfig {
+  fit?: 'content' | 'house';
+}
+```
+
+- no House Plan server-config or layout field is added;
+- no model/store/schema version or backend validation changes;
+- no data migration and no write to Home Assistant storage;
+- old cards and old YAML remain on the `content` path;
+- current code normalises only at read/render and visual-editor boundaries;
+- an older frontend ignores the unknown Lovelace key and renders its historical
+  content frame, without damaging House Plan data;
+- rollback is therefore data-free.
+
+## i18n
+
+Add complete keys to `src/i18n/{en,ru,de,fr}.json` for:
+
+- Framing;
+- All visible content;
+- Tight to house geometry.
+
+The technical YAML literals `content` and `house` are not translated. Missing
+translations may not leave a raw key or an unlabelled select option.
+
+## Performance and bundle
+
+- Frame resolution remains one bounded O(rooms + walls + openings + columns)
+  pass per structural render candidate.
+- HA state-only ticks reuse/calculate the same structural fingerprint; they may
+  not introduce DOM measurement or `getBBox()` layout reads.
+- No second hidden SVG, canvas raster scan, network request or dependency.
+- Existing `fit: content` cost and output remain unchanged.
+- Initial View bundle budget is not raised.
+
+## Touch and themes
+
+- `houseplan-space-card` is a View surface, so narrow phone and wall-tablet
+  rendering is release-blocking.
+- The new mode is presentation-only and must not cause horizontal overflow,
+  clipped structural strokes or a zero-height card at representative 320 px
+  and 900 px widths.
+- Pointer/tap behaviour is unchanged; the stage stays inert and the footer tap
+  target stays usable.
+- Light/dark and day-cycle modes share exactly the same structural frame.
+
+## Affected files and modules
+
+- `src/space-card.ts` — public `fit` type/default and static-render projection;
+- `src/space-editor.ts` — localised framing dropdown and legacy/default
+  projection;
+- `src/space-render.ts` — select ordinary vs structural frame and keep one `vb`;
+- a narrow pure helper in `src/space-geometry.ts`, `src/render/opening-symbol.ts`
+  or a dedicated static-frame module — finite structural items and opening
+  envelopes without DOM measurement;
+- `src/i18n/{en,ru,de,fr}.json` — editor labels;
+- `test/**` — pure frame/config/opening-envelope regressions;
+- `demo/smoke_space_card.mjs` — real-card comparison and View/touch contract;
+- `docs/USER-GUIDE.md`, `docs/USER-GUIDE.ru.md`, `docs/ARCHITECTURE.md`,
+  `docs/CANVAS.md` — public and canonical contracts;
+- `docs/CHANGELOG.md`, `docs/CHANGELOG.ru.md` — release bullet linked to #373.
+
+## Acceptance criteria
+
+- **AC1 — default compatibility:** missing/empty/unknown `fit` and explicit
+  `fit: content` produce the same numeric `viewBox`, stage aspect and #372
+  title interaction as the pre-change control fixture. **Evidence:** unit tests
+  plus `demo/smoke_space_card.mjs` comparing four configurations.
+- **AC2 — tight structural frame:** on a fixture with an oversized transformed
+  backdrop, far decor, far marker and room label, `fit: house` excludes those
+  items from its bounds, removes all four intentional 5% pads and yields a
+  strictly tighter frame than `content`. **Evidence:** pure unit assertions and
+  browser smoke over the rendered SVG `viewBox`.
+- **AC3 — no structural crop:** rooms, positive/zero walls, independent walls,
+  drafts, columns and door/window/gate envelopes in their extreme states are
+  all within the `house` frame including visible strokes. **Evidence:** pure
+  geometry table tests and a browser semantic-pixel/bounds witness that fails
+  when any structural edge is clipped.
+- **AC4 — stable and safe fallback:** an empty space, image-only space,
+  collinear structural geometry and invalid/unknown `fit` always produce a
+  finite positive frame; image-only `house` equals the ordinary content frame.
+  Opening/device state ticks do not change it. **Evidence:** unit + browser
+  smoke with state substitution.
+- **AC5 — editor/i18n contract:** the GUI displays the default and house choices
+  in en/ru/de/fr, saves exact literals, reloads them, and never offers cover.
+  **Evidence:** unit/source contract plus browser editor smoke or review of an
+  existing faithful `ha-form` harness.
+- **AC6 — View/touch parity:** tight mode renders without horizontal overflow
+  at 320 px and 900 px, in light/dark/day-cycle, keeps the stage inert and keeps
+  the footer action unchanged. **Evidence:** targeted browser smoke and code
+  review.
+- **AC7 — performance:** state-only rerenders do not perform DOM geometry reads
+  and the structural pass stays bounded; bundle budget remains green.
+  **Evidence:** code review, `npm run bundle:budget` and existing static-card
+  performance profile before beta.
+- **AC8 — documentation and release:** RU/EN guide documents `fit`, its default,
+  the intentional auxiliary-object crop and its interaction with `title: ""`;
+  architecture/canvas docs describe the second static-only frame; both
+  changelogs link #373. **Evidence:** docs/process gates and code review.
+
+## Automated test plan
+
+1. Pure unit: validate the `fit` resolver (`undefined`, `content`, `house`,
+   unknown) and prove the default branch retains current padding.
+2. Pure unit: build a synthetic structural scene containing a polygon room,
+   thick wall envelope, zero wall, independent wall/draft, column and every
+   opening type; compare exact finite structural bounds and zero padding.
+3. Pure unit: substitute closed/open/extreme opening amounts and prove the
+   state-independent envelope and resulting frame do not change.
+4. Pure unit: add far backdrop/decor/marker/label items; prove ordinary content
+   widens while house bounds do not. Add a detached valid room and prove it is
+   retained rather than outlier-rejected.
+5. Pure unit: empty/image-only/collinear/invalid inputs fall back to a finite
+   positive ordinary frame.
+6. Extend `demo/smoke_space_card.mjs` with control, explicit `content`, `house`
+   and `house + title:""` cards. Assert numeric frames, no clipped structural
+   edge, common overlay coordinates, unchanged footer deep link and inert stage.
+7. Repeat tight mode at narrow light and wide dark widths and change a door
+   sensor/device state; the viewBox remains stable and no overflow appears.
+8. In the implementation cycle run `npm run typecheck`, `npm test`,
+   `npm run build`, `npm run bundle:sync`, `npm run bundle:budget`,
+   `node scripts/no-new-any.mjs --base origin/dev --head HEAD`,
+   `node scripts/check-docs.mjs`, smoke selection and the named target smoke.
+   Run full golden/smoke/performance and exact-SHA Linux Validate before beta,
+   following the process.
+
+## Risks
+
+- **Zero padding clips the outer half of a wall or opening stroke.** Mitigated
+  by structural visible-envelope math and AC3, not by reintroducing arbitrary
+  padding.
+- **Backdrop/decor still votes through `spaceFrame()`.** Mitigated by a separate
+  explicit structural-item path and AC2; the ordinary helper remains untouched.
+- **Stateful door movement resizes the card.** Mitigated by one conservative
+  state-independent opening envelope and AC3/AC4.
+- **A detached building is mistaken for an outlier.** Mitigated by using all
+  sane structural items for the house union and the detached-room test.
+- **The default changes accidentally.** Mitigated by an explicit resolver whose
+  fallback is `content`, numeric compatibility tests and no implicit opt-in.
+- **Auxiliary content is intentionally clipped but undocumented.** Mitigated by
+  the GUI wording, guide warning and AC8.
+- **Static and full cards drift.** The new path is static-only by contract; all
+  existing shared geometry primitives stay canonical, while the full card's
+  content frame remains unchanged and covered by existing canvas tests.
+- **Docs screenshot fingerprint changes after `src/**`.** The canonical Docs
+  screenshots workflow/acceptance remains required by the repository process.
+
+## Rollback
+
+Remove the `fit` projection/editor option and always call the existing content
+frame path. Existing cards without the option are already on that path; cards
+with `fit: house` become ordinary content-framed cards when read by the rolled
+back frontend. No House Plan config/layout data, migration or cleanup is needed.
+
+## Release artifacts
+
+- `docs/CHANGELOG.md` and `docs/CHANGELOG.ru.md`: opt-in tight house frame with
+  a link to #373;
+- `docs/USER-GUIDE.md` and `docs/USER-GUIDE.ru.md`: YAML/table/default,
+  auxiliary crop and `title: ""` interaction;
+- `docs/ARCHITECTURE.md` and `docs/CANVAS.md`: static-only structural frame and
+  its relation to the canonical content frame;
+- targeted `demo/smoke_space_card.mjs` result with numeric and semantic visual
+  evidence for standard/tight, narrow/wide and light/dark;
+- full existing golden verify to prove no default-path visual regressions; a
+  new baseline is added only if targeted semantic evidence cannot make AC2/AC3
+  unambiguous, and may be accepted only from the reviewed complete Linux
+  artifact;
+- canonical Docs screenshots artifact accepted with
+  `npm run docs:accept -- --reviewed --from=<artifact>` because `src/**` changes;
+- no backend, migration or security artifact; initial bundle budget and the
+  static-card performance profile remain mandatory before beta.
+
+## Assumptions accepted provisionally; reviewer may change freely
+
+- the public literals are `fit: content | house`; names of internal helpers and
+  exact file placement are technical choices;
+- `content` is explicit current semantics rather than a newly invented mode;
+- “house geometry” means the complete structural list in contract item 5, not
+  room polygons alone, because room-only bounds would cut thick outer walls and
+  opening leaves;
+- objects deliberately excluded from bounds remain in DOM/SVG and may be
+  clipped; they are not hidden or deleted;
+- `house` falls back to `content` when no structure exists rather than framing
+  an arbitrary stored square or rendering an empty card;
+- a conservative analytic opening envelope is acceptable if it is tight to the
+  maximum rendered symbol and state-independent; DOM `getBBox()` is forbidden;
+- exact synthetic fixture names and sub-pixel tolerances belong to the test
+  implementation, provided a one-pixel crop causes a deterministic failure.