[AdaWorldAPI]

The reachability probe measures a partition and prints a table. That partition is a graph — addresses are nodes, calls and references are edges — and the tier that draws this exact wire already ships: medcare-rs emits MGRA v3 from graph_abi.rs, a2ui-graph renders it over wgpu with a WebGL2 fallback. What was missing was the emitter. No new renderer, no new format.

What landed

src/mgra.rs — the encoder. A pure writer, plus a GraphBuilder holding the address→index mapping the renderer needs (the wire carries node ordinals; a walk produces addresses). It refuses three things the wire cannot express: an endpoint past the node lane, a label over its u16 prefix, a count over the header u32.

The consumer is not defenceless — its edge_pairs() drops ghost ordinals so they cannot index a GPU buffer out of bounds — but a dropped edge is silent data loss: the field renders slightly wrong and nothing says so. Refusing the stream names the offending edge instead.

examples/mgra_emit.rs — the walk. Same recursive descent as the probe, with one extra piece of state: every queue entry carries the entry its descent started from, so a call becomes an edge rather than just a target. Kept separate from the probe's walk deliberately — the probe's numbers are pinned in a merged PR, and rewriting its walk to serve a second consumer is how a measured result quietly moves.

The falsifier

The round-trip test decodes with a2ui-graph's own parser, added as a dev-dependency with default-features = false. Zero wgpu crates enter the lock; 30 transitive ones do (ndarray, the wasm-bindgen/web-sys family). An encoder checked against your own re-read of your own encoder proves only that you were consistent.

It earned its keep immediately — it caught the first of the two mistakes below.

Three things that went wrong, and what they taught

1. edge()'s third element is the predicate @11; the line style is a separate byte @8 behind edge_kind(). The first version wrote the relation into @8 alone. A call and a data reference differ both in meaning and in how they should be drawn, so both fields carry it.

2. The first walk observed a node per decoded instruction — nodes == instrs exactly (1349/1349, 7347/7347). That is an instruction dump, not a call graph.

3. The data arm fired zero times, and rather than assume why, the address operands were counted by space:

   space	count	in image
   Unique	1257	—
   Register	354	—
   Const	166	none

   6502 absolute addressing lifts its address into a computed temp, so recovering an in-image data address needs constant propagation through the p-code temps — which this walk does not do and does not fake. What Const does catch is the out-of-image absolute, and the emitted addresses are exactly what that predicts: $0007, $0057, $0059, $005b — zero page.

   (in_image is deliberately not applied to a Unique operand: its offset is a temp slot number, not an address. 92 of them compare "inside the image" by coincidence of numbering.)

   A guard now warns when either arm reports zero — that column read 0 and looked like a measurement.

A finding this surfaced elsewhere, not fixed here

reachability_probe's own data_refs set is structurally empty on these images, for exactly the same reason. Its unreached+ref column is not measuring what its name says.

Left alone on purpose: its numbers are pinned in a merged PR and in CLAUDE.md, and quietly rewriting a measured result is how one moves. It wants its own change with the re-run as the evidence.

Measured

image	entries	referenced	calls	bytes
LOCODE.unprot.bin	79	4	40	2757
HICODE.unprot.bin	199	10	526	12783

Artifacts are gitignored on the same footing as the ore — a build product derived from a commercial game's image.

Gates

cargo clippy --features lift --all-targets -- -D warnings clean · cargo fmt --check clean · cargo test -p ruff_r2il --lib 46/46 · prek on every changed file, all hooks pass · both probes re-run end to end.

---

Generated by Claude Code

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 4ca2ce2e-04bd-45b8-b601-83dcda2fe724

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: serverGenReqId_7b48e6ec-208d-4346-ae54-5aeb19f8d835)