git:20260906.e43b2e0 to git:20260906.0839e8a

+ 0b. **Purpose, asked before the contract in step 0 (MANUAL).** Ask once, at task start: *render-only / 3D print / both?* — recommendation **both**. Record it in `builds/<slug>/state.md` as `purpose: <answer> (owner <time>)` and in `design-parameters.json`. No answer → `purpose: assumed both` and build print-ready; a later downgrade is noted, never a rebuild. Purpose already stated in the request → do not ask. Purpose fixes the mesh strategy before the first pass:

+

+ | Purpose | Solids & booleans | Walls / holes / keys | Curved-surface segments | Scale |

+ |---|---|---|---|---|

+ | `render-only` | closed enough for the camera; overlap may stand in for a boolean | relief may be shader-only; blind holes acceptable | as low as the silhouette tolerates | 1 BU = 1 m |

+ | `print` / `both` (default) | every part one closed solid, signed volume > 0, transforms applied, EXACT boolean + `checkpoint()` — never faked by overlap | real walls ≥ spec `min_wall_mm`, holes modelled through, keyed flats modelled (`keyed_flat_mm`) | ≥ 64 (default, editable per part) on any surface the gate measures a diameter or wall on | 1 BU = 1 m, spec in mm |

+

+ Getting this wrong costs a rebuild, not a tweak: `render-only` geometry cannot be gated later without re-modelling the walls.

+

+ 3. **Critic (self-check):** are import/`__file__` inside the payload (the MCP namespace does not persist)? have socket/enum/operator names been introspected at runtime? units (1 BU = 1 m, spec in mm)? mutation before assert? destructive op → `checkpoint()` first? does the mesh match the 0b row for the recorded purpose (closed solids, real walls, segment count)?

+ 5b. **Two gates when purpose is `print` or `both` (MANUAL).**

+ - **Form gate — the closing condition of phase 2 (build phases: 1 blockout · 2 form + interfaces · 3 detail + materials · 4 studio + composition · 5 delivery), before any material, studio or delivery-media work:** temporary identity bake of the printable parts (`pass-NN-print-prep` pattern) → `spec.json` written from measurement (skill `blender-agent-core` → recipes "Spec from measurement") → `production-gate.py` exit 0. **No *delivery* still, film or turntable starts before this gate is PASS or waived.** *Delivery media* = the images/films/turntables produced for the owner in phases 4–5. **Verification imagery is exempt and stays MANDATORY:** verify-ladder rungs 2–6 (`preview_render`, viewport screenshot, low-sample Cycles, comparison sheet, turntable) and per-pass clay/pose renders are how a pass proves itself (see step 5 and §Visual feedback) — they are never blocked by a gate. Verification imagery stays cheap: ≤ 512 px or low-sample, ≤ ~2 min GPU per pass; anything costlier, or anything the owner is asked to accept, is delivery media and needs the gate. Gate ≈ 3 s; the delivery media it protects ≈ 1 h (2026-09-06 watch winder: a 0.9 mm knife edge surfaced only after 16 stills and 2 films — 53 min GPU re-rendered).

+ - **Final gate — at delivery**, on the delivered print bake. Both gates must PASS or be waived; a report bound to another scene/spec sha is not reusable.

+ - **Gate waiver:** one owner-quoted line in `state.md` — `GATE WAIVED <time> "<owner words>"`. README manufacture then reads NOT_REQUESTED. Silence is never a waiver.

+ - Gate fails → `refine-spec` (root cause in the spec) or `refine-code`; max 3 runs, then `request-input`.

+ - Purpose `render-only`, or no printable part → both gates are skipped and README manufacture = NOT_REQUESTED. Say so explicitly; do not leave it blank.

+ | Gate `wall_thickness_screen` fails at ~1 mm on a lathe part that is 2.5 mm thick everywhere | a bore/opening meets a curved inner wall → knife edge at the lip (watch winder shell: 0.9 mm) | truncate the cavity with a flat ceiling one wall below the rim; a D-keyed hole needs `keyed_flat_mm` + `keyed_flat_dir` in the spec, not a wider `tol_mm` |

+ | Boolean cuts a hole with no walls / leaves boundary edges, or `DIFFERENCE` empties the target | target or cutter mesh is inside-out (signed volume < 0 — procedural lathes/fans are the usual culprit) | assert signed volume > 0 and 0 non-contiguous edges on BOTH operands before adding the modifier (`builds/watch-winder-capsule/scripts/ww_mesh.py::orient_outward`) |

+

+ **Media evidence void (geometry changed after media exists, MANUAL):** never re-render automatically. The README media rows carry `rendered from <blend sha>` and a `current <blend sha>` line marked *pre-change*; then ask the owner **exactly one** question, with a GPU estimate per media set (stills / film / turntable), and quote the answer in `state.md`. Owner declines → the mismatch line stays in the README permanently, that is the honest record. Owner accepts → archive the old set (`renders/<set>-pre-<tag>/`) and never overwrite it.