E0597 (27 sites, measured d72ffe8708, same producer/subject as the E0004 lane): 100% ONE rustc message - '<X> does not live long enough'. Not a mixed population: every single site is the same borrow-lifetime shape. Concentrated in v2_compiler_translate.rs=10, then v2_compiler_parse.rs=4, v2_compiler

[briansrls]

---

Enrolled evidence (added after review)

The board measurement above proves the repair works today; nothing in the tree would notice it regressing tomorrow. DESIGN §4b(4) keeps a class's discriminating RED and its positive control enrolled, and a measurement living only in a PR body is not enrolled. So: dag/test/claim/emitter_callable_field_capture_witness_test.dag, four rows.

Under dag/, not src/v1/tests/claim/ — CI invokes --source-root dag --source-root src/v2, so a witness under src/v1 is discovered by nothing.

It carries both failure modes, because this repair has two. A witness that could not tell them apart would pass for the wrong fix. The preamble text the assertions require is unproducible by the pre-fix shape (no move) and unproducible by the naive-move shape (no preamble), so one assertion refuses both. The fourth row asserts init does not clone scale: a capture set widened to every local in scope would be correct at every site, unbounded in cost, and indistinguishable from a precise one by any assertion that only checks the build is green — that row makes the derivation executable, not merely the correctness.

Executed on both binaries, not reasoned:

row	pre-fix binary	fixed binary
arrow_typed_field_lambda_owns_its_captures	false	true
sibling_arrow_field_captures_every_local_it_names	false	true
derived_capture_set_excludes_a_local_the_lambda_does_not_name	true	true
non_callable_field_takes_the_unchanged_path	true	true

A direct probe asserting the pre-fix text init: Rc::new(| returned true on the pre-fix binary — the RED is the emitter's actual output, not an inference. Three arms are additionally pinned at rustc: pre-fix shape → E0597 ×3, naive-move shape → E0382, landed shape → clean.

Correction to a claim this PR body previously implied

An earlier reading reported unreachable_pattern as measured-and-zero across both arms. That was wrong and is withdrawn. curated_cargo_probe_one.sh builds its coded column from ^error\[E[0-9]+\] and its uncoded column only from lines beginning error: ; a warning: unreachable pattern line matches neither, so the instrument cannot emit that class. The nonzero I had cited to make the zero readable (uncoded_unreachable_pattern:21) comes from a 2026-07-26 TSV written by a different vintage of this producer — a control borrowed across frames.

unreachable_pattern is therefore unmeasured at both arms, and no weaker claim is put in its place. The histogram claim above is unaffected: it was always about coded errors only, and it stands — exactly one line of difference, E0597:27 removed, nothing added.

Why no dispatch was spent measuring it: every changed byte is confined to a record-field value position, adding no match arms and reordering none, so this repair cannot make a sibling arm unreachable. That is an argument from the diff's shape, labelled as such, not a measurement.

The same argument covers the head move for the witness commit: a new module under dag/test/claim/ that nothing imports cannot enter the src/v2/compiler/03_ingest.dag closure the probe measures, and that commit changes no emitter byte — so the one-line histogram diff is preserved by construction.

The grain of the "resolved, not moved" claim

Stated precisely, because class counts and site identities are different evidence and the stronger-sounding reading is not the one measured here. What is measured is at class grain: across the two arms, E0597 drops by 27 and no other class rises by any amount — every remaining coded row is identical and both totals fall by exactly 27. That rules out this repair trading its class for another one, which is the failure mode "resolved vs moved" names.

What it does not do by itself is join diagnostics on (emitted file, line, col), so it cannot distinguish "these 27 sites are gone" from "these 27 are gone and 0 unrelated sites appeared" — the two are indistinguishable when the second number is zero, which is why the claim survives here, but the grain is worth naming rather than leaving implied. A site-identity join would upgrade it; nothing in this PR rests on the upgrade.