svg: rasterization moves below the exporter; the water preset declares its absorption

[krazyjakee]

Two changes an engine consuming the lowered SceneGraph needs, both of which
also remove a reach-around Studio was already doing.

Requested by a downstream consumer (moggy, krazyjakee/moggy#457), but neither
change is engine-specific — both are about the built-in surfaces being
reachable and self-describing on their own terms.

1. mogen-svg: the rasterizer moves below the exporter

Rasterization shipped as a pre-export pass (#106). That is still the right
shape for rewriting a SceneGraph's texture paths and layering bytes over a
TextureSource, and it is unchanged here. What was in the wrong place was the
renderer underneath it.

Every consumer holding a Material and a path sits beside the exporter, not
after it. svg.rs's own module docs already named the problem — "the consumer
this pass cannot reach" — and Studio's viewport solved it by reaching back into
mogen-export for is_svg / render_svg / resolve_svg_size. An engine that
reads the lowered SceneGraph and decodes textures itself cannot do even that
without linking the whole exporter, oxipng / fbxcel / meshopt included,
for three pure functions over (bytes, size, wrap).

So the policy moves into a new mogen-svg crate — mogen-core plus
resvg/usvg, both pure Rust — and the pass stays in mogen-export, which
re-exports the four public items. mogen_export::is_svg and friends still
resolve, so Studio needs no edit. textures-svg becomes dep:mogen-svg.

There is still exactly one implementation of "what pixels does this SVG
produce", which was the point of the original decision; it is simply reachable
now. Verified to cross-compile to wasm32-unknown-unknown, as the old module
claimed for itself.

2. The built-in water preset declares absorption

shader="water" was a bare on/off switch: the preset was seeded with an empty
param list, so an author could not say how absorbing the water is, and a
consumer reading the lowered Material found a shader name and nothing else.

It now declares absorption (float, default 1.5), resolved through the
ordinary ShaderDecl::resolve_param path — shader_params (absorption=2.5)
overrides it exactly as for a user-declared shader. The built-in stays the
first client of the general system rather than acquiring a special case, which
is the property #107 is heading toward anyway.

The default is deliberately not zero: zero is the "behaves exactly like clear
glass" case, so defaulting to it would make shader="water" a no-op that
reports success. The test asserts the override wins, that a silent material
gets the default, and — the control — that the default is not the disabled
value.

docs/dsl.md documents the parameter in both places water is described.

Verification

• cargo test --locked -p mogen-svg -p mogen-export -p mogen-dsl -p mogen-core — 18 suites, all green (664 in mogen-dsl, 27 in svg_textures)
• cargo check --workspace --all-targets — clean
• cargo check -p mogen-svg --target wasm32-unknown-unknown — clean

Not run: the Studio GL paths, which need a display seat this machine may not
have. Studio is untouched by construction — it compiles against the same
re-exported names — but that is an argument, not an observed run.

🤖 Generated with Claude Code

[krazyjakee]

Reviewed this PR against code quality, file structure, test coverage, and a visual check of the rendering path it touches. One real gap found and fixed; everything else checks out.

Findings (ranked by severity)

1. Low — missing integration test for the new absorption param actually reaching the export (fixed)

The PR's new lowering test (crates/mogen-dsl/src/lower/tests/shader.rs) only checks that ShaderDecl::resolve_param resolves the right value on the SceneGraph. Nothing exercised the actual deliverable this feature is for — per the PR description, "a consumer reading the lowered Material" (i.e. a downstream engine) needs the value in the exported glTF's node.extras.shader.params, and there was no test pinning that, unlike the existing declared_shader_params_merge_defaults_with_overrides coverage for user-declared shaders.

Fixed in dccebe9: added builtin_water_shader_projects_absorption_default_and_override to crates/mogen-export/tests/shader.rs, asserting both the default (1.5) and an authored override (2.5) land in the exported extras. Also manually verified end-to-end through the real mogen CLI (release build, not just cargo test):

// node.extras for shader_params (absorption=2.5)
{ "shader": { "name": "water", "params": { "absorption": 2.5 } } }

2. Informational, no action needed — Studio's live preview doesn't consume the new param yet

crates/mogen-studio/src/viewer/flatten.rs:586 and user_shaders.rs:101 still hard-code the built-in water branch to shader id 1 with an empty param list, so shader_params (absorption=...) has no visible effect in Studio's GL viewport today. This is a pre-existing, explicitly tracked limitation (user_shaders.rs cites #107 by name), not a regression introduced by this PR — the data-model-level work here is a reasonable prerequisite for that follow-up. Flagging only so it's not mistaken for a bug if someone tries shader_params (absorption=…) in Studio and sees no change.

Verification performed

• cargo check --workspace --all-targets --locked — clean.
• cargo test --locked -p mogen-svg -p mogen-export -p mogen-dsl -p mogen-core — all green, including the new test.
• Confirmed the mogen-export::svg → mogen-svg crate split is behavior-preserving: built examples/features/svg_texture.mog through the actual release mogen binary and extracted the embedded rasterized textures from the produced GLB. dots.svg (wrap/tile mode) and emblem.svg (fit mode) both rasterize cleanly with correct colors and seamless wrap-around, no artifacts.
• Confirmed feature wiring (textures-svg = ["textures", "dep:mogen-svg"]) and re-exports (mogen_export::{is_svg, render_svg, resolve_svg_size, MAX_SVG_SIZE}) are intact — mogen-studio/src/viewer/gl_util.rs compiles unchanged against them, matching the PR's "Studio needs no edit" claim.
• No file touched by this PR exceeds 1000 lines; no split needed.

Nice PR — the crate split is clean and the doc comments carry real signal.

---

Generated by Claude Code