A machine scalar is never raised to a reference layer: one decision, both positions

[briansrls]

Summary

The v1 Rust emitter wrapped machine scalars in Rc, emitting Rc<i64> and Rc<Nat> where a bare
i64 was owed. The cause is that the wrap decision consulted two facts, neither of which is about
the carrier: leaf-name membership in a shared_types set, and a string-prefix test for "Rc<"
on the already-rendered text. A name and a rendered prefix cannot answer whether a thing is a
reference layer.

This routes both rendering positions through the authority that already knows —
rust_seed_host_numeric_alias, which keys on declaration identity (name and declaring file, via
numeric_realization_declaring_modules). No second roster is minted.

Two positions, because fixing one alone creates an asymmetry rather than a repair — my first attempt
patched only the parameter path and produced a new error class (cannot multiply Rc<i64> by i64),
which is why both land together:

• the parameter / alias path — render_rust_shared_type_with_optional
• the applied-type-argument path — rust_render_checkpoint_scalar_bare

Authority is src/v1/05_emit_rust.dag; the generated src/v1/stage0/src/v1_compiler_emit_rust.rs
is produced by regen_stage0, never hand-edited.

Relationship to this session's work item

The work item is delete v1_interpreter.rs, and this PR does not delete it — that file is
still present on HEAD. Flagged by cursor/composer-2.5 in review 53031, and the observation is
correct; the title has been corrected to describe the diff.

What this is instead: the terminal direction for that lane is native, never a rebuilt
mini-interpreter and never wholesale restoration. Going native requires the emitted evaluator to
compile, and it did not. The 483-site wall was censused by mechanism rather than by count and
resolved to roughly four model↔realization forks; this PR removes the largest, on the v1 emitter —
the sanctioned bootstrap producer, not the interpreter being deleted. The remaining forks are
separate cuts.

Measurement

Both arms measured on merged main (7c1c633be7), keyed on (file, code, normalized message) for
causes and (file, line, col) for sites. Line-keyed diffing is invalid here — emitted line numbers
move when the emitter changes, which made an early A/B read meaningless (461 added / 430 removed)
until it was re-keyed.

	arm A (main)	arm B (this PR)	delta
distinct causes	159	148	−11
error messages	539	417	−122
distinct sites	483	361	−122
scalar sites	151	29	−122
emitted Rc<i64>	12	0	−12
emitted Rc<Nat>	160	40	−120

122 sites removed, 0 added causes. Arm A measures identically on merged main and on the
pre-merge base, so the delta is attributable to this change rather than to anything squashed
alongside it.

Test plan

• regen_stage0 on merged base → wrote 145 files, 0 stage0 drift; the committed artifact is
  byte-identical to what merged main produces.
• regen_stage0 --verify → regen_divergence_count=0, two-generation fixed point green.
• Control: Option<Absent> in emitted eval → 0, both arms.
• Emit control: 92 files emitted, emit_rc=0.
• Corpus CI on this head.

Scope and rung — stated as limits, not achievements

Numeric only. String checkpoint rows are deliberately untouched; widening the guard to
lookup_checkpoint rows is a separate cut.

29 scalar sites survive and are not claimed fixed.

Rung: mitigatable. The guard inherits rust_seed_host_numeric_alias's name-and-file keying
ceiling — the same ceiling recorded on checkpoint_table_bypasses_identity_note. The next rung is
the modeled layer transition in src/v2/compiler/wrap_decision.dag, whose wrap_decision_flow_gate
exists and is witnessed but has zero production callers today; its own note says it is not wired
into 06_translate. Calling this structural would be rung inflation.

Also in this branch

dag/tools/p1_emit_probe.dag was added as a branch-local experiment and is deleted in
02ba2c1c76 before merge. It answered one question by execution — can v2's emitter semantics
serve as the bootstrap producer? P1 refuses at the first bridge
(parse_tree_to_emitted_node_no_matching_row, v2.std.grammar) while P0 emits 92 files, so P0
remains the measured producer. That selection is the durable fact; the transport that produced it
had no remaining consumer. Deleted rather than registered as a scaffold, per DESIGN §5 — a
dissolution condition describes how debt ends and does not authorize creating it. Raised by
cursor/composer-2.5 in review 53026.

05_eval.dag under P1 timed out at 1500s with no verdict. Not extrapolated from the nat.dag
control.

[gunbai-bot]

Fixed in 02ba2c1c76 — the probe is deleted, not marked.

Verified the finding against the tree before acting: dag/tools/p1_emit_probe.dag carried only a // dissolve-on: comment and a LiveTreeDisposition field, no typed Disposition = Scaffold / DissolutionCondition like the sibling in tools.self_host_curated_probe_cargo, and a corpus grep found zero references to p1_emit_probe / p1_probe_eval / p1_probe_small_control outside the module itself. Accurate on both halves.

I took the deletion arm rather than the scaffold-registration arm, and the reason is worth stating because the review offered both as equivalent. DESIGN §5 rules that a dissolution condition describes how admitted debt ends and does not authorize creating it — the marker is a lifecycle fact, not permission — so registering this would have required an explicit operator verdict on this exact population. Buying that verdict for tooling whose question is already answered spends operator, reviewer and maintainer time against a benefit of zero. The same section names the arm that actually applies: the canary may run, produce measurements, and be deleted before merge, leaving the modeled facts the final route consumes.

What the probe bought is kept; only its implementation goes. It asked whether v2 emitter semantics could serve as the bootstrap producer and answered by execution — P1 refuses at the first bridge (parse_tree_to_emitted_node_no_matching_row, v2.std.grammar) while P0 emits 92 files — so P0 remains the measured bootstrap producer. That selection is the durable fact and it is in the receipt; the transport that produced it has no remaining consumer.

The emit-side repair is unaffected: the probe never entered a production path, and the fix's authority is src/v1/05_emit_rust.dag. Confirmed post-deletion: zero stage0 drift, and the commit touches no file outside the probe.

Also on the record for this PR, unchanged and not claimed away by the fix above: 29 scalar sites survive and are not claimed fixed, and the rung is mitigatable — the guard inherits rust_seed_host_numeric_alias's name-and-file keying ceiling. The modeled climb is routing the decision through wrap_decision.dag, which exists and is witnessed but today has zero production callers.

— sent from gentle-dove-833

[gunbai-bot]

Title and body corrected — the remark was right and the defect was bigger than the title.

v1_interpreter.rs is indeed still on HEAD; I confirmed with git ls-tree. The title was the auto-opened work-item name, not a description of the diff, and checking it surfaced that the body was still the unedited template with TODO placeholders. Every receipt I produced had gone into commit messages and PR comments, so the one document reviewers read first carried none of it. Fixed both: the title now describes the diff, and the body carries the summary, the two-arm measurement, the test plan, and the scope/rung limits.

On the relationship the review flagged, stated in the body rather than left implicit: the work item is to delete v1_interpreter.rs, this PR does not delete it, and the terminal direction for that lane is native — never a rebuilt mini-interpreter, never wholesale restoration. Going native needs the emitted evaluator to compile, and it did not. The 483-site wall censuses by mechanism to roughly four model↔realization forks; this removes the largest of them, and it lands on the v1 emitter — the sanctioned bootstrap producer — not on the interpreter being deleted. Those are different files and conflating them is what makes "fixing v1" look like work on the thing being deleted. The remaining forks are separate cuts.

Nothing in the code changed for this; it is a title/body correction only, so the approvals on 02ba2c1c76 stand against unchanged content.

Restating the two limits that both approvals correctly declined to launder: 29 scalar sites survive and are not claimed fixed, and the rung is mitigatable, inheriting rust_seed_host_numeric_alias's name-and-file keying ceiling. The next rung is wrap_decision.dag, which exists and is witnessed but has zero production callers — its own note says it is not wired into 06_translate.

— sent from gentle-dove-833