Module shape_registry 

Process-global record-shape registry (#462 slice 3).

Maps sorted field-name lists to stable u32 shape IDs for records built at runtime — JSON decode, SQL row, HTTP body, host effect handlers, builtin returns. Anywhere Value::record_dynamic is called.

Why: the slice-2b measurement (docs/design/ic-polymorphism-measurement.md) found 14% of GetField IC hits — and 100% of inbox/gateway/std_http traffic — landed on records carrying NO_SHAPE_ID. Those records all aliased on the sentinel, so the IC’s shape-keyed verification (added in #517) couldn’t distinguish them and fell through to the name-compare path on every hit. With this registry, dynamic records built from the same field set share a real shape_id and the IC hits cleanly amongst them.

Disjoint ID spaces

Compile-time Op::MakeRecord shape IDs come from Program::record_shapes and are per-program indices (0, 1, 2, …). Runtime registry IDs come from this module and live in the high half of the u32 range starting at DYNAMIC_SHAPE_ID_BASE. NO_SHAPE_ID (u32::MAX) is reserved.

Disjoint spaces are required for IC correctness: the shape-keyed verifier in vm.rs treats cached_shape == incoming_shape as “offset is sound” for non-NO_SHAPE_ID cases. If a compile-time ID collided with a dynamic ID for a different field set, the IC would return values from the wrong field.

Cost

One RwLock<IndexMap> lookup per record_dynamic call. The map grows monotonically (shapes are never freed) but is small in practice — tens of distinct shapes per workload. Lookups are read-lock + hash; misses take the write lock once per new shape.

Sorting

The key is the sorted field-name vec — two records with the same fields in different insertion order share a shape_id. Matches the existing Value::Record PartialEq (which compares IndexMap contents structurally, ignoring shape_id), so equality and IC sharing line up.

Constants

DYNAMIC_SHAPE_ID_BASE

Base for dynamically-interned shape IDs. Chosen so that Program::record_shapes indices (which start at 0 and grow up) cannot collide with dynamic IDs even for pathologically large programs. Compile-time records will hit a different IC slot outcome than dynamic records with the same field set — that’s the cost of keeping ID spaces disjoint; the slice-2b measurement found 0 mixed-flavor sites in real workloads.

Functions

intern

Look up (or assign) the shape ID for the given field-name list. Same set of names — in any insertion order — yields the same ID for the lifetime of the process.