The emitted-text attestation reads code outside string literals, not prose

[gunbai-bot]

The emitted-text attestation reads code outside string literals, not prose

Defect. v1.compiler.emit_rust reference_derived_use_line_plan admits a proposed use-line only when the name appears among the emitted source's identifier tokens. That gate exists to reject misproposed candidates. But the token set was built from the whole emitted text, string-literal contents included. std.content_hash's grounding note mentions v2.std.node.Hash. The emitter proposed Hash for std.content_hash, and v2.std.node declares type Hash = Fnv1a64Structural, an alias of std.content_hash's own type. The prose then let that candidate through, and std_content_hash.rs emitted pub use crate::v2_std_node::{Hash};. Nothing uses it. It is harmless in one crate but is a reverse edge in any partition, because v2.std.node imports std.content_hash.

How it was found. The Pkg7e emitted-workspace plan (v2.workflow.emitted_crate_workspace, on its own branch), once #12190 and #12200 made the two closures agree, refused at the edge merge with the typed EmitterReverseEdge std_content_hash -> v2_std_node. That refusal was added in #12147.

Earliest unjustified link (DESIGN §6b): the gate treating string contents as emitted code. rust_code_outside_string_literals removes escaped backslashes and quotes, splits on the remaining quote characters, and keeps the segments outside literals. The token set is built from that. The candidate producers are untouched, and emit_typed_expr and render_rust_type are untouched, per the operator ruling.

Evidence, by execution.

• Discriminating specimen: gunbc compile --entry a probe module importing v2.std.node { Node }. std_content_hash.rs carries the v2_std_node use-line before the change (1 line) and not after (0).
• Whole seed (claim_executor --required-regen):
  • Generation 1, emitted by the old binary, drifts only in the emitter's own mirror.
  • Generation 2, emitted with the fix active, drops eleven use-lines from six mirrors: v1_compiler_emit, emit_rust, infer, trait_derive_emit, gunbc_occurrence_binding_parser_walk, and one claim module. Each is a name that appears in its module only inside a string literal, e.g. AlgebraFieldTemplate and OccurrenceId in v1_compiler_infer.rs.
  • The seed compiles without them, which proves they were dead.
  • Generation 3 is a fixed point (first_generation_equal=true).
• Failure mode rostered: gunbc.recurring_failure_mode a_string_literal_attests_an_emitted_reference. Rung: mitigatable. Ceiling: structurally guaranteed. Next-rung trigger: use-lines derived from the emitted reference set the renderers return, the capability emitted_inline_crate_path_edges_unreturned_stall names.

Second arm, same gate. Over the full v2.compiler.compile closure, std_content_hash.rs still emitted the Hash use-line. Its code carries #[derive(..., std::hash::Hash)], and the tokenizer splits on :, so a segment inside a qualified path attested the bare name. rust_qualified_segments_marked prefixes every segment after :: with @, so a marked segment never equals a candidate. A path head stays a token, so Parent::Variant still attests Parent. Seed regen: the next generations drop the remaining re-exports whose names occur only as qualified segments (in emit_rust, infer, infer_resolve, parse and trait_derive_emit). The seed compiles, and the third generation is a fixed point.

Limits, stated. The stripper reads quote characters, so a '"' char literal would flip its parity. If that happened, it would drop tokens, and missing imports fail loudly at rustc; it cannot silently add a use-line. The whole-seed fixed point is the measured evidence that no such case exists in the seed today.

v1 admission. This is v2 self-host work under gunbc.v1_maintenance_standing v1_seed_standing: the emitted use-lines are the edges that decide whether the emitted compiler partitions into crates. It adds one function to the emitter .dag and one call-site change. The mirror changes are regenerated, not hand-edited.

🤖 Generated with Claude Code