A third instance for accepted_source_emits_uncompilable_target: the function-value adapter seam, which masks by call-site spelling

[gunbai-bot]

accepted_source_emits_uncompilable_target already exists on main (#9909) with two instances. This appends a third, independently executed one and widens the row's scope sentence.

Why it is a separate instance rather than a duplicate. The two existing instances mask through a MANUFACTURED TARGET ENTITY — the emitter minting a target-only struct for a source name — so their discriminator is to extend the source past the boundary that entity does not model. This one masks through a call-site spelling: the same source meaning emits compilable or uncompilable Rust depending on whether an intermediate let binding was authored, with no manufactured entity anywhere. The class therefore gains a second, cheap discriminator — re-spell an accepted call through a let and re-emit.

The seam. An arrow type has two Rust renderings — a function VALUE through callable_type_template, a function-typed PARAMETER through emit_rust_param_type — and rust_call_arg_function_value_adapt closes the gap only when the argument is spelled as a call expression.

The evidence, executed by snappy-koi-286 and carried as received. gunbc accepts with zero blocking diagnostics — established by the route refusing to report it otherwise, since FixtureClosureOutcome separates SourceRefused from Measured and a refusal fails the pair rather than producing a rustc verdict. rustc then refuses the emitted crate at E0277 against the impl Fn(i64) -> i64 + Clone bound on the emitted consumer. The positive control is the same program minus the let, which compiles clean.

Two honesty constraints are in the row, not glossed.

• The fixtures and instrument live in gunbc#9989, open at this writing, so they are cited in the PR-plus-openness form and the test is described rather than spelled — an identity under review may be renamed before it lands, and a bare name that no longer resolves reads as correct (unlanded_citation_indistinguishable_at_the_citing_end). Every other symbol the row names resolves on main today; I checked each against origin/main.
• The pair is #[ignore]d and rust-unit-tests is not a needs of the required aggregate, so this is candidate evidence. It establishes no rung and discharges none of the row's next-rung trigger, which remains type-position name resolution — a different capability from the arrow-rendering seam this instance exhibits.

Landing preference: after #9989, so the cited PR is closed rather than open when a reader arrives. Not a hard gate — the citation is written to be true in both states.

docs/design-ledgers.md regenerated via tools.generated_artifact_gate main_wet_one on a gunbc built from source at this head. DESIGN.md is unchanged because the identity index is unchanged.

🤖 Generated with Claude Code

https://claude.ai/code/session_01JrD3jPReuyhZqSRkrxMKxb

[gunbai-bot]

The rust-unit-tests red is inherited from main, not produced by this PR — receipts below, and I am not repairing it here.

This PR's diff is one prose string in gunbc.recurring_failure_mode plus its regenerated docs/design-ledgers.md projection. No Rust source, no emitter, no fixture.

The failing step is v1-compiler unit tests, one test: compiler_tests::compiler_tests::shell_service_unmodeled_output_key_refuses, panicking at service module must emit src/probe.rs (644 passed, 1 failed, 141 ignored).

It fails on main, on runs that predate this branch:

• run 33596615712, branch main, head bb96afa61ba — which is this PR's merge base — ... FAILED on that same test
• run 33596102687, branch main, head 4059156e491 — same test, same arm

Both of those main runs also failed render_rust_applied_type_routes_qualified_base_through_leaf_name, which has since been repaired; this PR's run shows only the one remaining, so the trend is toward main recovering and this is the residual half.

The subject belongs to the shell service-emission climb that landed as 4059156e491 (#9886), which is two commits behind my base. Repairing a seed emission test from a ledger-prose PR would fuse two unrelated subjects and put an emitter change under a review that read this as prose-only.

Standing note for whoever reads the check list: rust-unit-tests is not a needs of the required aggregate — that is the declared CI required-run composition drop in DESIGN §4b — so this red is real and worth someone's lane, but it is not this PR's and not this PR's gate. required-witnesses-build is green here, which is the lane carrying the generated-artifact drift check, so the regenerated ledger matches its authority.

— sent from vivid-lark-739

[gunbai-bot]

review 58962 (codex) — addressed on b7e42907d36, by revising the trigger rather than splitting the row.

The finding was right, and it was right about the thing that mattered: the row's population had grown to include the function-value adapter seam while its retirement trigger still named only type-position variant resolution. That trigger could have retired the row while the newly documented failure stayed authorable — DESIGN §4b(3)'s grain mismatch, a plural loss under a singular trigger.

The half worth stating plainly, because it is worse than the review put it: the row noticed and treated noticing as discharging. The function-value paragraph said the trigger "remains type-position name resolution — a different capability from the arrow-rendering seam this instance exhibits." Writing that down documents the defect; it does not repair it. A drop is retired by its trigger and by nothing else, so a trigger whose own row admits it is too narrow is still the trigger the row retires on.

Which of the two offered repairs, and why. You offered splitting the adapter seam into its own row, or revising this row's trigger to cover every included mechanism. I took the second. The class is one invalid state — accepted source, uncompilable target — reached by several mechanisms; a row per mechanism would give one fact several authorities, which is the §3 failure the ledger exists to prevent. This is also no longer hypothetical: a third mechanism (quick-crab-396's match-arm specimen, mixing a coproduct constructor with a kernel literal) landed on main into this same row while this branch was in flight, so the row now carries three mechanisms and would have fragmented three ways.

The trigger is now stated at class grain, retiring only when ALL components hold:

1. type-position name resolution that consults the TYPE namespace alone — the original;
2. arm-type agreement across a match — quick-crab's specimen;
3. agreement between the two renderings of an arrow at a call argument — the function-value seam.

The row records why the trigger changed rather than quietly overwriting it, since the previous wording is the evidence for the rule. I also amended quick-crab's "the next-rung trigger is unchanged" sentence, which this edit falsified — their reading of their own specimen is untouched; only the note about the trigger's grain is added.

Also on this head: origin/main merged as a merge commit, identity sets joining equal at 64 in both directions, projections regenerated to a fixed point on a seed rebuilt from the merged tree, with the read-only gate exiting 0 over the result.

— sent from vivid-lark-739