Record the passive-edge mask, so multi-edge ops can be emitted

[0-jake-0]

Stacked on #761 → #760 → #759. Base is eligibility; this diff is only the top commit. Review order: #759 → #760 → #761 → this.

What this changes

NodeInfo records the op's passive-edge mask and exposes edges_in_call_order(). The emission walker consequently handles join and sample instead of refusing them — emission now covers single-edge, multi-edge and passive-edge shapes.

Why

Gap 3, the last of the three #759's spike pinned.

join only needed the receiver-first convention active_ups already keeps. sample needed something the graph did not have. Its data leg is edge 0 and passive, its trigger is edge 1 and active, so the node records:

active_ups  = [tick]
passive_ups = [count]

That pair is identical whichever way round the call was written. The two lists are a partition — each preserves its own order, but the interleaving between them is gone, so count.sample(&tick) and a hypothetical reverse are indistinguishable.

The op's #[op(passive = [..])] mask is exactly the missing bit, and the derive already had it. Recorded on the node now, with edges_in_call_order() walking positions 0..n and taking from whichever list each bit selects.

How it was verified

• cargo fmt --all
• cargo lint ✅ — cargo lint-all not run: --all-features cannot build here (aeron needs CMake ≥3.30, box has 3.28; etcd needs protoc). Documented prerequisites, unrelated to this diff; CI is the first real check on that feature set.
• --all-features blocked as above. Full default suite plus bench,async,csv,cache,augurs,market,ws,redis,postgres,kafka,web,kdb,prometheus,fix,dynamic-graph,tracing — green.
• New behaviour covered by tests asserting values and tick times

Each shape is proven the same way as the rest of the spike: the expected artifact is a real nitro! block in the test file, so byte-identical output plus a compiling file means the emitted wiring is valid by construction. Parity then runs on values and tick times.

The sample test asserts the reconstructed order explicitly ([data, trigger]) as well as the emitted text, because that is the case where getting it backwards would still produce something that compiles — n0.sample(&n1) is valid Rust that computes the wrong thing.

Notes for the reviewer

Variadic ops stay refused, and correctly so. merge_all → MergeN and combine → CombineN have hand-written forwarders, so they carry no build name. Emitting them as the n-ary call they do not accept would be worse than failing, since it would produce an artifact that fails at pass 2 in generated code. A test pins the refusal rather than leaving it to chance.

What remains after this: refusals name node indices rather than the call sites #[track_caller] would give; tier-2 captures (func!([thresh] |x| …)) are unbuilt; and the walker is still test-only code rather than a codegen::generate entry point.

Stack note, repeated from #761: this is now four PRs deep on unreviewed work, and #759 is load-bearing for the three above it. If #759 wants changes, the rebases compound. The remaining #726 items are mutually independent, so this is a natural place to stop adding layers.

---

Generated by Claude Code

[0-jake-0]

Superseded by #769, which squashes this whole stack (#759–#768) onto main as a single reviewable change.

The passive_mask and edges_in_call_order work is carried forward in full. Closing to keep the queue readable; no content is being dropped.

---

Generated by Claude Code