music-video-gen/flow-state/MIGRATION.md

Migrating a scene onto the identity artifacts

The recipe tools/migration-status.js reports progress against. Sixty-one scenes is too many to convert from taste, so this is written to be followed mechanically and to produce the same result twice.

Background is EPIC-3.md. The one-line version: a scene draws its own content, a migrated scene draws the song's content — its cast, in its ink, on its lattice — and the measured payoff is additive with everything the scene already did.

node tools/migration-status.js          # the queue, and what each scene needs
node tools/migration-status.js --all    # every scene and its tier

---

The rule that decides everything

An artifact is content a scene could not have invented for itself.

If the scene still looks right with the identity switched off, it has used the artifact as a modifier and the migration has not happened. That is not a style preference — it is the exact failure mode of u_sigSides, which thirty scenes declare and most quietly ignore, and it is why the gate below exists.

---

The four tiers

The classifier assigns these from the shader source. Check its answer, don't trust it: it reads patterns, and a scene that places elements in an unusual way will be misfiled.

tier	what it looks like	consumes
drawn	loops over discrete elements at computed positions	cast, ink, staging
figure	one or a few forms, its own composition	cast, ink
field	a continuous surface — noise, flow, terrain	ink
treatment	an effect over an image rather than an image	ink
solid	a camera in a space, marching a distance field	form, ink

Accents are excluded. They are mostly-empty depth passes, not subjects.

---

Recipe: drawn

The highest payoff and the least invention. Four edits.

1. Declare it. Add to the module, above params:

consumes: ['cast', 'ink', 'staging'],

2. Replace the primitive with the cast. Whatever the loop was drawing — a circle, a box, sigShape, a bespoke SDF — becomes one of:

float d = castMain(q) * size;      // the protagonist: large, few
float d = castChorus(q) * size;    // the chorus: small, many

q is the element-local coordinate, (p - pos) / size. Multiplying the result back by size restores it to scene units, which is what the ink expects.

Use castMain when the scene draws a handful of things and castChorus when it draws a crowd. A scene that draws both should use both — that is what the two members are for.

3. Replace the placement with the lattice.

vec3 node = stageNode(fi, float(u_count));   // xy position, z scale multiplier
vec2 pos = node.xy;
float size = u_size * node.z;

Keep whatever the scene did that was motion — a drift, an orbit, a wander, a march. Give up what was composition. The split is the point: identity owns where things are, the scene owns what they do.

A scene whose composition IS its identity — a spiral that must be a spiral — can keep it and take stageScale() alone, which is the song's element size. Declare staging either way.

4. Replace the edge with the ink.

col = mix(col, pal(i + 1), inkMask(d, uv));

inkMask does fill, fill treatment (hatch, stipple, halftone), outline and edge hardness in one call. Delete the scene's own smoothstep(soft, -soft, d) and its sigEdge — the ink supersedes both.

Then wrap the return:

return vec4(inkValue(col), 1.0);

Do not delete sigCamera, sigGrain, sigHorizonY or sigAir. Traits and artifacts are different layers and both still apply.

---

Recipe: figure

Steps 1, 2 and 4, skipping the lattice. Declare consumes: ['cast', 'ink']. Take stageScale() if the figure has a size worth scaling and add staging if you do.

---

Recipe: solid — the cast in three dimensions

For a scene that has a camera in a space rather than a plane: a raymarcher, a corridor, anything where the subject can be walked around. Declare consumes: ['form', ...].

vec3 ro = vec3(0.0, lift, -dist);          // eye, in object radii
vec3 rd = normalize(fw * lens + rt * p.x + up * p.y);
vec3 n;
float hit = castMarch(ro, rd, dist + 3.0, n);
if (hit > 0.0) col = castLit(n, rd);       // lit in the track's palette

castSDF3(vec3) is the distance field if you want to place, repeat or carve with it yourself; castNormal3 is its normal. All of them fall back to the flat profile extruded when a track brought no assembly, so they are safe to call unconditionally.

Many instances

For a scene that was stamping castMain in a loop — a field, a swarm, a belt — swap the stamp for castSolid, which marches one instance orthographically in its own frame:

vec2 local = (p - centre) / size;          // exactly what castMain was given
if (dot(local, local) > 1.6 || painted > 0.5) continue;
vec3 n;
float hit = castSolid(local, castTurn(yaw, pitch), n);
if (hit > 0.0) { painted = 1.0; col = castLit(n, vec3(0.0, 0.0, 1.0)); }

castChorusSolid is the same for a chorus member — the protagonist's body plan with fewer parts and its own proportions, which is what a field of many should be drawing.

Both guards in that snippet are load-bearing, and each was found by a measurement rather than by review:

• the bounding-sphere reject, because without it every pixel evaluates every instance's distance field — Swarm measured 59ms/frame at 4K against a 60ms ceiling;
• painted, because instances overlap several deep at the top of the size range, and marching all of them made Floating Geometry's own gate run for minutes. Which instance wins where they overlap was always arbitrary, so first-wins costs nothing.

A third rule lives in the contract rather than in your scene: take the surface normal AFTER the march loop, never inside it. GLSL unrolls a fixed-bound loop, so a normal in the loop body multiplies four more copies of the assembly SDF by the step count. For the same reason these helpers are compiled only into scenes that declare form — see FORM_PREAMBLE.

Worth the cost only if the shot MOVES relative to the object. A solid held at one angle is a silhouette with shading, and cast draws that for a fraction of the price — the assembly earns its keep through the outline changing, which needs either the object turning or the camera travelling.

---

Recipe: field and treatment

There are no elements to replace, so this is one edit plus a judgement.

consumes: ['ink'],

return vec4(inkValue(col), 1.0);

If the field already dithers, hatches or posterises internally, replace that with inkPattern(uv) so the treatment is the song's rather than the scene's. If it does not, inkValue alone is the whole migration — a value structure shared across every scene in a video is worth having and costs one line.

Be honest about treatment scenes. Most of them are effects wearing a scene's clothes, and EPIC-3 §8 argues they should move into the identity's EFFECTS register rather than compete for screen time as subjects. Migrating one is a holding action, not the answer.

---

Verifying — the part that makes this reproducible

A migration is not done when the code looks right. It is done when the gate passes:

checks.html?scene=<Scene%20Name>

Every artifact in consumes must show a passing line:

PASS  consumes: cast       delta 255/255 (floor 24)
PASS  consumes: ink        delta 255/255 (floor 24)
PASS  consumes: staging    delta 255/255 (floor 24)

That check renders the scene twice under two deliberately distant identities and requires the picture to change. A scene that declares cast and ignores it fails here, which is the only reason consumes can be trusted at library scale.

npm run lint:scenes enforces the other half in both directions: declaring an artifact without calling it, and calling one without declaring it. The second matters more than it looks — an undeclared artifact hides the scene from this report and from anything that later selects scenes on capability.

Then the usual battery still applies. renders something, animates, deterministic, distinct, param sweep, flash rate and every declared trait must all still pass. A migration that breaks distinct has made the scene into one of its neighbours, which is a real risk here: the more scenes share a cast, the more two weakly-composed ones converge.

---

Measuring the payoff

Per scene, the gate. Across the library, two numbers:

checks.html?decompose=1    identity against container, measured apart
checks.html?experiment=1   stages against legacy, with error bars

decompose is the one to watch. Before the migration, over the four purpose- built stages:

identity only    0.0299      container only   0.0557      identity = 54% of container

After the first eighteen scenes, over the eight that consumed the cast:

identity only    0.1287      container only   0.0815      identity = 158% of container

Complete, over all thirty-two:

identity only    0.1054      container only   0.0901      identity = 117% of container

The number that matters in that row is container only, which went 0.0557 → 0.0815 → 0.0901 as the migration progressed. That is the stop condition below holding: the library did not homogenise. identity only is noisier than it looks because the scene it probes is picked from the migrated set and changes between runs — Constellation, then Metaballs, then Floating Geometry — so read it as "comparable to the container" rather than as a precise ratio.

End to end, song variety across twelve songs — and a warning about how to read it. A single run after the migration gave separation 0.38. Three runs give:

separation   0.314, 0.555, 0.065      mean 0.311, half-range +/-0.245
floor        0.1073 +/-0.0145
observed     0.1162 +/-0.0185
ceiling      0.1434 +/-0.0000

Do not quote the separation figure. It is a ratio of two small differences — (observed - floor) / (ceiling - floor) — and the numerator here is 0.0089 while each of its terms carries an error bar twice that size. The ratio is not measuring the generator at this sample size, it is amplifying the noise in both. The 0.38 that appeared in a single run was meaningless, as was reporting it.

What can be said: the ceiling now sits reliably ABOVE the floor, with zero variance across runs, where before the epic it landed underneath and the ratio was not computable at all. That is a change in kind and it is solid. The magnitude is not.

The trustworthy measurement in this harness is the DECOMPOSITION, not the top-line ratio. decompose has a noise floor of exactly 0.0000 and effects an order of magnitude larger than any drift, because it compares renders directly rather than dividing differences of aggregates. Steer by that.

Coupling did not move: -0.02 +/-0.03. Whether a song LOOKS different in proportion to how it SOUNDS different remains unsolved, and nothing in the migration addressed it.

Read that with its caveat: the stage set changed underneath the measurement, so part of the jump is that Metaballs expresses a cast more strongly than Constellation did rather than that the migration itself moved anything. What it does establish is the thing worth knowing — a migrated LIBRARY scene carries the identity better than a stage written from scratch to carry it. The bespoke stages were the wrong shape of effort. As scenes migrate, the identity only number should climb while container only holds — because the whole point of the correction in EPIC-3 §9b is that these add rather than trade. If container only falls as scenes migrate, the migration is homogenising the library and should stop.

---

A caution learned the expensive way

This harness's noise floor is large enough to invent findings. Three claims in this epic were made from single runs and withdrawn after repeats: that roster size was the dominant lever, that stages beat legacy by 14%, and that the identity's range was the bottleneck. Anything under about 0.01 of spread needs &repeats=3 before it is believed, and a difference that changes sign with the sample size is not a difference.

Migrate in batches, measure after each batch, and expect the per-batch effect to be inside the noise. The trend across batches is the signal.