Context-aware assets

You are furnishing a scene whose facts TEE already knows: plan dimensions,
GPS datum, site-photo palette, and the client's spoken brief. Never guess
what these tools can tell you, and never pull raw catalogs into context —
state WHAT you need and let the server rank, gate, and validate.

Hard rules (they are enforced server-side; work with them, not around):

• Licenses are gated fail-closed. If an import answers license_blocked,
  pick another hit — there is no override and you must not ask for one.
• Uniform scaling only; doors/windows/appliances/sanitary/seating/humans
  never stretch. A reject from the scale policy means the asset is
  wrong, not the policy.
• Code-severity placement rules (door swings, egress widths, toilet
  clearances) never relax. Guideline rules relax only via an explicit
  relax: ["rule_id"] on the plan item, which is recorded.
• Text before pixels: verify geometrically (as_verify); render at most
  ONE budgeted image per task, only when the report says
  render_warranted.

The 7-step checklist

Follow in order; each step names its exact tool.

1. Brief — as_style_brief() for the palette (named colors), style
   terms, and avoid-list. tee_recall() for project conventions. Note the
   GPS datum and region (US/EU rules) from the extract facts
   (ex_search("gps")).
2. Search — one as_search(query=…, asset_class=…, max_tris=…, match_style=true) per furniture class you need. State target
   dimensions from the plan (e.g. the plan's 0.9 m door opening) — do not
   browse. ≤5 rows come back; pick by dims first, tags second. If two
   candidates tie, as_sheet/tee_media is the tie-breaker, not the
   default.
3. Fit — decide target_dims per asset from plan facts (openings,
   wall lengths). Reference: reference/dimensions.md for the class
   envelopes you may aim inside.
4. Plan — write ONE relational placement plan for the whole room:
   [{name, class, dims, anchor: "<wall-id>", offset}, …] (~10
   tokens/object). Anchor wall-backed classes (sofa, bed, wardrobe) to
   walls; leave circulation to the validator.
5. Validate — as_place(plan=…, room=…, region=…) WITHOUT apply
   first. Fix every violation by editing the plan (the fix field says
   how); relax a guideline rule only with a reason you state in your
   reply.
6. Apply — imports first (as_import with target_dims and
   location from the solved plan), then as_place(apply=true) with the
   entity ids. Then materials: as_material(id, query) for measured PBR
   (lane 0) or as_photo_material when the surface exists in site
   photos. Sun last: as_sun(lat, lon, when, apply=true).
7. Verify — as_verify(room=…, match_style=true). Fix violations and
   re-verify. Only if clean AND a visual question remains, ONE
   tee_capture(max_kb=32). Finish with as_credits() so attribution
   ships with the project.

Judgment room

Selection within a shortlist, grouping (a reading corner vs scattered
chairs), and style calls are yours — the tables bound dimensions and
clearances, not taste. When the brief's avoid-list conflicts with a
search hit (e.g. "no marble"), drop the hit even if it ranks first.

References (one level deep)

• reference/dimensions.md — class envelopes + fit-to-plan targets
• reference/style-matching.md — palette/ΔE00 ranking, avoid-lists
• reference/lighting.md — sun facts, HDRI bands, world rotation
• reference/sources-licenses.md — backend license/ToS matrix
• evals/scenarios.md — the three acceptance scenarios for this skill