Single source of truth for capture value semantics (inference vs emission)

[zharinov]

Problem

Two independent implementations answer "what value does this capture hold":

• crates/plotnik-compiler/src/analyze/type_check/infer.rs — flow-based, produces the declared types (infer, .d.ts generation).
• crates/plotnik-compiler/src/compile/capture.rs (creates_structured_scope and friends) — syntactic, decides what the bytecode actually emits.

When they disagree, release builds produce type-unsound JSON. The only cross-check is debug_verify_type in crates/plotnik-vm/src/engine/verify.rs — debug builds only, at user runtime. Confirmed divergences, each is a corpus case:

1. Duplicate tagged-alternation labels silently overwrite in a BTreeMap; check passes, exec panics (verified 2026-06-12):

   cargo run -p plotnik -- exec -q 'Q = (program (expression_statement [A: (identifier) @x  A: (number) @y]))' -s 'foo;' -l javascript

2. Duplicate capture names inside a named node are accepted (or_insert; the sequence path correctly errors) → output is invalid JSON with duplicate keys (verified 2026-06-12):

   cargo run -p plotnik -- exec -q 'Q = (program (expression_statement (binary_expression (identifier) @x (identifier) @x)))' -s 'a + b' -l javascript
   # → {"x":…,"x":…}

3. Tagged alternation under a node capture: inference promises a tagged union, runtime emits the node (debug panic).
4. Uncaptured recursive refs are typed Void ("opaque"), but the VM bubbles their captures into the enclosing scope → duplicate keys.
5. :: string / :: TypeName annotations on array captures rewrite struct elements or discard the array shape → panic. The annotation logic is scattered across ad-hoc cases.
6. field: (Def) @cap infers flat, emits nested → panic.
7. Null vs optional: typegen renders x?: T, the materializer always emits "x": null — under strictNullChecks real output violates every generated .d.ts. Also * and + differ on missing arrays ([] vs null), contradicting docs/type-system.md.
8. Cross-file workspace inference re-walks a file with a fixed source_id, so refs into other files attribute diagnostics to the wrong source → out-of-bounds panic in the renderer.

Approach

• One function answering "what value shape does this capture hold" (MatchedNode | StructuredScope | RefResult | Array), consumed by both infer.rs and capture.rs.
• One apply_annotation(shape, annotation) that explicitly transforms, names, or rejects each combination.
• Per-def (not per-file) inference, with source attribution derived from the owning symbol.
• Decide the null-vs-optional policy once and align typegen, materializer, and docs/type-system.md.
• Items 1 and 2 are also missing validation (duplicate labels / duplicate captures should be diagnostics) — add those regardless of the unification.