W21B Wave 4 Hybrid Spatial Reference Implementation Brief

0. Locked Decision

0.1 Non-negotiables

• No scenario-specific production logic. L05/L09/L04 stay in docs, fixtures, and canary evidence only.
• No regex / substring heuristic as semantic SOT. IDs, enums, schemas, coordinates, and LLM/VLM structured fields are the contracts.
• No detailed pretty FP art. The scaffold is simple boxes, lines, icons, and metadata sidecars.
• No destructive asset cleanup. Stale ImageAsset / PNG history remains preserved.
• No BG or scene image canary until coordinate scaffold and shot-conditioned projection artifacts pass their evidence gates.

1. Current Touchpoints

2. Phase Plan

2.1 Phase A — Path-S coordinate scaffold

Add a coordinate layout plan that turns FP metadata into simple validated geometry. This is a data contract, not a rendered image contract.

2.2 Phase B — deterministic SVG scaffold renderer

Render validated layout JSON into simple SVG and optional raster. It must be byte-stable and API-free. It is not a place-richness illustration.

• Draw boxes, lines, doors, windows, and minimal fixture icons.
• Render each render_on_plan=true marker once.
• Expose metadata-only elements in sidecar/debug tables, not SVG clutter.
• Output: floor_plan_svg, optional floor_plan_svg_raster, and HTML review artifact.

2.3 Phase C — shot-conditioned projection sidecar

Generate camera_view_grid_10x10 only after selected-shot camera/look-at is known. Projection is not an FP-level generic property.

{
  "bg_id": "LxxByy",
  "fp_id": "fp_x",
  "selected_shot_id": "Sxx_Shotyy",
  "projection_timing": "after_selected_shot_camera_decision",
  "coordinate_sot_source": "path_s_layout_json|path_v_vlm_readback|hybrid",
  "projection_source": "computed_from_layout|vlm_derived|hybrid",
  "camera_request_source": "shot_aware_bg_render_plan|floor_plan_prompt_camera_recommendation",
  "camera_view_grid_10x10": [[{"markers": [], "summary": "", "confidence": 0.0}]],
  "visible_elements": [],
  "place_context": {}
}

2.4 Phase D — BG adapter consumption

Background prompt/render should consume the scaffold and projection sidecar instead of asking image T2I to infer detailed FP layout from a pretty plan.

• Add background prompt pack version only after A/B/C artifacts are stable.
• Prompt receives simple scaffold summary, shot-conditioned grid, and place-density tags.
• FP/SVG image attachment is optional support, not the sole layout SOT.
• BG visual canary stays active-only; no stale asset cleanup.

3. Commit Plan

4. Canary Plan

4.1 Dry-run sequence

1. Run Path-S layout plan on active target FPs. Text LLM only.
2. Render deterministic SVG previews. No image model.
3. Run shot-conditioned projection for representative selected BGs. Computed path first, VLM fallback within cap.
4. Emit summary JSON and side-by-side HTML: layout JSON, SVG preview, projection grid, evidence, conflict flags.

4.2 BG canary sequence

1. Only after dry-run passes, force background prompt/render for active BG subset.
2. Compare against v10/v11 baseline with active-only gallery.
3. Reject if BG ignores scaffold/grid, loses place richness, or relies on detailed FP art to recover layout.

5. Acceptance

5.1 Stop conditions

• Path-S requires scenario-specific literals to pass validator.
• Path-V returns evidence-poor grids below confidence threshold on simple probes.
• Projection is attempted before selected-shot camera/look-at is available.
• BG adapter reintroduces detailed FP art dependency.
• Implementation needs destructive asset deletion or manual per-row active edits.

6. Work Split