Add R40 Artifact Control + R41 Minimal Effective Edit — clean renders and safe edits

[aievolutionpl]

What this adds

Two rules that together cover the full artifact story: R40 for artifacts you generate, R41 for the ones you edit into existence.

Artifacting is reframed from bad luck into eight named failure modes with eight named causes, in two families. The fixes deliberately do not transfer between modes — re-rolling a style collision will never clear it, and sharpening a prompt will never clear context bleed.

Generation artifacts (R40)

Mode	Cause	Fix
A Texture dissolution — Voronoi cells, webbing	fine repeating detail asked for at scale	bound the detail
B Style collision — muddy, matches neither	two styles that can't coexist	one artist, one medium
C Context bleed — ghosts of earlier images	in-chat editing memory firing unintentionally	one concept per session
D Quality-tier grime	draft settings on a final asset	draft cheap, ship high

Edit artifacts (R41)

Mode	Cause	Fix
E Reference drift — the room/face/layout changed too	no preservation contract; silence reads as permission	write PRESERVE
F Edge halo / texture bleed	the edit was too large to integrate	shrink the ask, hold the lighting
G Duplication / warped geometry	the model re-solved a protected region	name geometry as protected
H Overload — no hero left	several ideas requested at once	one hero, one message

R41 — the part that matters most in production

Most commercial image work is editing, not generation. Its dominant failure is not a dirty texture — it is the model rebuilding what nobody asked it to touch:

	Asked for	Got back
❌	"replace the stove"	a redesigned room, moved furniture, new flooring
❌	"change the clothing"	a different face
✅	"replace the stove"	the same room, one new stove, matched shadows

• Two slots, never one: EDIT INSTRUCTIONS and PRESERVE. An edit prompt with only the first is an open licence to redraw the frame.
• Copy-paste preservation blocks for scenes/interiors, faces and products, plus an integration block (scale, perspective, contact shadows, reflections — an object with no contact shadow reads as a sticker).
• One element per turn. Modes E–H are all one failure at bottom — too much asked at once — so the meta-remedy is to make the ask smaller. Two clean edits beat one ambitious one.
• Reference accuracy outranks creative improvement. An edit that made the picture nicer while losing the room is a failed edit, not a bonus.

R40 mode C, corrected in the second commit

The first pass said "one image per fresh session." That was too blunt — it forbade the legitimate case of iterating on the image you just made. The accurate rule is one concept = one image family = one session: direct iterations (angle, lighting, framing, one object, a text fix) stay in the session; a new concept, scene or campaign direction opens a fresh one.

The failure people actually hit is running a new family inside an old one — and different subject is not protection, since unrelated prompts ghost each other just as readily.

Also worth flagging

• The high-risk subject register mapped onto this repo's real briefs — food (seeds, crumb, char), fashion (knit, weave), hotel (foliage, stonework), pets (fur), jewellery. Not exotic: roughly half our briefs need the pre-flight.
• Constraint discipline — use only the negatives a scene can actually produce. Every possible constraint in one prompt dilutes the ones that matter; no point banning malformed hands in a packshot with no people.
• Reduce micro-detail, don't just ban it — "a small number of subtle natural sparks", not "thousands of glowing sparks". Quantity is what dissolves into noise, not the element.
• A symptom→remedy table covering all eight modes, so a dirty output is diagnosed rather than re-rolled.

Wiring

• SKILL.md — core rules 23 and 24; step 4 gains the artifact pre-flight and the edit branch; step 5 inspects at 100% and diffs edits against their reference.
• visual-advertising-engine.md — 39 → 41 rules.
• references/qa-gate.md — the vision pass inspects at full resolution and returns texture_artifacts, unbriefed_elements, style_coherence, edge_halo_or_bleed, duplicated_or_warped and unrequested_changes; four new hard fails (R40-texture, R40-ghosting, R41-drift, R41-halo), each routed to the mode that causes it rather than a generic "regenerate".
• references/prompt-library.md — a reference-image edit template built on the preservation contract, plus the anti-artifact constraint block.
• anti-slop-registry.md (new §1a, "the slop nobody chose"), model-routing.md, design-rules.md and core.md, so the standalone chat inject stays self-contained.

Why scripts/qa.py is untouched

A deterministic artifact check was built and rejected during this work, not skipped.

Sensor grain, fabric weave and Voronoi webbing are all the same high-frequency energy to a script. A busy-block metric flagged a legitimate high-ISO photograph at 1.00 while scoring synthetic Voronoi webbing at 0.00 — an exact inversion. An autocorrelation variant scored the repo's own canonical photo+panel layout (1.00) as worse than pure noise (0.003).

Any threshold strict enough to catch webbing rejects real photography, so shipping it would have made the gate worse. Artifacting is called by eye at 100% in layer 2, where the distinction is actually decidable — and the QA gate now says so explicitly rather than leaving a silent gap.

Verification

python scripts/check_docs.py → PASS (30 markdown files: links, §-pointers, rule IDs, frontmatter, version)
python scripts/test_qa.py → 13/13 cases passed

Purely additive: R01–R39 unchanged, no rule-ID renumbering. Version 5.6.0 → 5.8.0 across SKILL.md, both READMEs, CHANGELOG.md and plugin.json (which had drifted at 5.5.0 and is now realigned).

Note on sources

apipass.dev — the URL in the original request — is blocked by this environment's network egress proxy, as are rewarx.com and aigcsdm.com. The R40 material was reconstructed from search-result summaries of those articles, which carried the substance (the four categories, the ghosting mechanism and its "first image is clean" pattern, the Impressionist + ultra-detailed collision example, the verbatim negative-constraint string, the layout-specification framing, the quality-tier guidance). R41 comes from the skill text supplied directly in the conversation. Worth a spot-check against the original article for any R40 specific I could not read firsthand.