fix(schema): make the exported schema accept what the engine accepts

[LeadcodeDev]

rustmotion schema is what an editor and a generating model read to know what to write, and it rejected every one of the fifteen scenarios this repository ships as examples. Draft-7, 62 failing paths. All fifteen validate now, and the one-letter typo that used to produce a 3611-node error tree reports in 10 nodes at the property that is wrong.

#424 named two causes and measured both. Measuring the rest turned up two more, so this is four fixes rather than two.

The templating vocabulary. A templated scenario writes "delay": "$delay" where the schema declares a number; the engine substitutes before it deserializes, so the file is valid input. Every position declaring a number, an integer or a boolean now also accepts ^\$[A-Za-z_][A-Za-z0-9_]*$, through a single TemplatePlaceholder definition — inlining it at the several thousand sites cost 690KB where the $ref costs 310KB. A node carrying an enum is deliberately left alone: that keeps type a closed set, which is the discriminator the second half depends on. And the pattern means "delay": "soon" is still an error, rather than one more thing the schema waves through.

div's five aliases. The schema did not know card, flex, grid, container or positioned, which the examples use throughout. every_serde_alias_in_the_sources_is_reachable_from_the_exported_schema was green the whole time: its scanner kept a source line only if that line both started with #[serde( and carried an alias = ", and rustfmt splits a six-argument attribute across six lines. progress_bar fits on one line and was caught, which is what made the test look like it worked. The scanner now takes each attribute whole, paren-matched across lines, and the reachability check knows a variant alias widens an enum rather than adding a sibling property.

Two hand-written deserializers describing their Rust shape. These were not in the issue and were 52 of the 62. AnimationTiming parses through AnimationTimingWire: loop was declared required although the wire defaults it, declared a boolean although it also takes a play count (#330), and repeat_count was advertised as a property although it is an output of RepeatSpec::into_parts and the wire type is deny_unknown_fields — believing the schema there cost you the whole component. FontWeight's visitor takes "normal", "bold" or a number; the derive emitted "Normal", "Bold" and {"Weight": n}. AnimationTiming derives neither Serialize nor Deserialize, so its serde attributes are read by schemars alone and changing them cannot move runtime behaviour. Six more types have the same shape and are unverified — #462.

Steering. Every branch already pinned its tag to a one-element enum, so the tag was never what was missing. A union whose branches pin one key to disjoint values becomes an allOf of if/then on that key: validation is unchanged, and a validator descends into the branch that was meant instead of failing all 54 and reporting whichever it liked. The component/for-each/use union has no common tag and is steered on key presence. fade_in_upp now reports at scenes/0/children/0/style/animation/0/name; a for-each without its template says so; propz instead of props is named.

The decision #424 left open — whether the schema should validate a templated file or only an expanded one — is settled here in favour of templated. The README sells the schema for editor autocompletion and LLM prompts, four of the fifteen examples are templated, and the skill tells a generator to factorise.

boon joins as a dev-dependency, for the test that compiles what the command actually prints and validates every example against it. It locks four new packages against jsonschema's seventy-seven, and cargo audit with CI's ignore list reports nothing new.

Verification — cargo fmt --all --check, cargo clippy --workspace --all-targets --features rustmotion/studio -- -D warnings, cargo test --workspace --features rustmotion/studio: 0 failed. The example test was confirmed red with the widening removed. The 62 → 0 figure was measured twice through different validators, Python jsonschema 4.26 while working and boon 0.6 in the committed test.

Closes #424
Refs #462