Image → 3D (v3 — Blender native)

Build from supplied references without new generated/retrieved vendor meshes, hosted model calls, paid credits, or third-party downloads. Load blender-agent-core first and use references/review-rubric.md for visual verdicts.

Phase -1 — Fidelity contract

Record before construction:

1. Target use: hero still, turntable, animation, realtime, or print.
2. Observable fidelity target and critical features in each supplied view.
3. Treatment of unseen surfaces: mirrored/plausible inference or more references required.
4. Production constraints: editability, topology, materials, dimensions, and output views.

For a mechanical or industrial reference, designate one primary view for visible count, facing, placement, and silhouette; use close-ups for local section and construction. Before choosing functional routing, ports, instrument stations, or connection anatomy, read the applicable project drawings or official installation/technical documents. Record whether each rule is a fact, a design choice, or unknown. CAD can support a profile or architecture only when its match and units are recorded; it never establishes identity, dimensions, selection, or compliance by itself.

For a multi-component site, declare one common scale and shared units/axes. Record sourced physical envelopes separately from display envelopes, protrusions and service allowance; mark proxy/analogue dimensions provisional. Name external ports with their elevation, orientation and role, and verify the modules together before final integration. An illustrative scale does not establish equipment selection or capacity.

Do not promise exact hidden geometry from insufficient evidence. Visible reference fidelity has priority.

Pin local reference bytes with source roles before comparison. An original photo, an owner-supplied dimension, a generated blueprint/360 sheet and a manufacturer drawing have different authority. Preserve conflicts and label local adaptations; later generated views cannot authenticate unseen switch, shaft or PCB geometry. When only inline images are visible, disclose missing byte provenance and obtain the real file rather than inventing a local path or hash.

Phase 0 — Native reconstruction mode

If references cannot constrain the requested fidelity, stop at the contract and request the missing views. External generation/retrieval is not a fallback.

Phase 1 — Evidence specification

• Measure normalized extents and landmark ratios from every reference.
• List critical visible components; every item maps to a named Blender object or surface region.
• Define scene graph, symmetry, units, material groups, cameras, and render outputs.
• Store the plan under plans/; build idempotently under builds/<slug>/.

Phase 2 — Staged construction

1. Blockout: primary silhouette, proportions, camera match.
2. Form: shared surfaces, duct/body transitions, bevel hierarchy, panel boundaries.
3. Detail: functional parts, apertures, fasteners, ribbing, seams.
4. Materials: local Blender node materials only; neutral diagnostic lighting first.
5. Polish: production naming, transforms, topology, render settings, outputs.

Use data API/bmesh first. Render and file-save operators are allowed where necessary. Keep components separate when editability or animation requires it.

For a new geometry family, the controller first proves the identity-risk prototype at the target apparent scale. Promote a constructor only after its explicit axes, pivot, ports, accepted parameter range, and regression proof have passed for a second real consumer. Treat an instrument body and its process adapter as separate contracts; a camera-visible head is not placement evidence.

Phase 2.5 — Offline retopology gate (frozen scaffold only)

QRemeshify 1.1.0 may be used only when the user selected an already-local scaffold. First create a manifold bounded input with scripts/prepare-qremeshify-input.py; then run scripts/run-qremeshify.py. The wrapper must pass pinned-library hashes, macOS quarantine approval, a 1k–100k triangle input budget, and a hard wall timeout. For highly fragmented or fused scan/generated scaffolds, start without sharp guidance; dense boundary features can crash the native field stage.

Run tests/blender/qremeshify-candidate-test.py, restore/project materials if needed, then render hero and top comparison sheets with scripts/render-qremeshify-candidate.py. Accept the candidate only when topology improves without dimension, silhouette, aperture, propeller, or feature loss. Otherwise keep the Blender-native repaired scaffold. Prefer retopology before components are fused; whole-object retopology of a monolithic drone can smooth or tear small mechanics even when numeric topology passes. QRemeshify changes topology; it does not repair inaccurate reference geometry.

Phase 3 — Verify every pass

Run numeric checks before image checks. Then generate a camera-matched comparison sheet:

scripts/make-comparison-sheet.sh <reference> <render> <sheet>

Choose one verdict: continue, refine-spec, refine-code, request-input, or stop. Two repeats of the same visible failure require changing the modeling approach, not another parameter tweak.

For every declared critical feature, review the primary and supporting views manually at a comparable apparent scale. Name the feature, expected construction layers, source/object/interface, proof shot, and falsifier. A numeric count, clean topology, or an unreadable distant object is not a visual pass. Check risky bores, coplanar joins, fastener families, lettering, dial marks, and specular edges at the final output density before the full batch. Preserve source, camera/settings, and media hashes; a changed source, camera, or resolution invalidates affected proof.

The HQ coverage checker can validate declared receipt bindings with or without --launch. Its guarded --launch route validates first and then controls only its child render. Direct Blender, MCP, and headless renders bypass that launch guard; record the route accurately and do not imply a universal pre-HQ interception.

For Full HD/social delivery, load knowledge/60-pipeline/native-render-delivery.md. Render the actual saved model. Preflight every shot at final aspect, check curved metal surfaces and legends at delivery density, and inspect original images plus decoded animation. A concept-generation image, thumbnail or enlarged screenshot does not qualify a native Blender render. Changes to normals/materials still invalidate affected shading and media proofs even when vertices are unchanged.

Delivery gate

For repeated reference refinement, use the target-bound critic packet in knowledge/60-pipeline/native-agent-iteration.md. Freeze the target/reference, candidate sources and actual proof captures; ask a separate reviewer for one PASS/FAIL/UNKNOWN observation and actionable correction per contracted feature. scripts/native-review.py checks the declared evidence bindings and missing findings. It does not judge images itself or authenticate reviewer independence. A repeated failed feature calls for a different approach; changing the target starts a new contract revision. The controlled native coupon illustrates a missing detail that passes overall dimensions; it is not a general reconstruction benchmark.

Separate contracted feature acceptance from exact likeness. Retain omitted legends, camera differences, RGB tuning and inferred underside details as visible deltas. Missing critical evidence remains UNKNOWN; never weaken the rubric after seeing a failed candidate. A completed still review does not approve unreviewed motion or the physical/electrical design.

• Deterministic rebuild succeeds twice with stable counts.
• Critical components, transforms, topology, materials, camera framing, and render settings pass automated tests.
• Hero and secondary reference sheets are inspected at output resolution.
• Opposite, underside, and turntable renders expose no missing surfaces or camera failures.
• Report remaining visible deltas and inferred surfaces explicitly; do not inflate fidelity.
• If QRemeshify was used, record version, library hashes, settings, elapsed time, topology metrics, comparison verdict, and fallback result.

No live vendor link may remain in delivery. A user-selected local scaffold must be appended as a copy, repaired into a self-contained asset, provenance-tagged, and verified with external-service-calls = 0.