[kixelated]

Problem

lang was one crate of 31.7k lines with one case list, one caller.ts, and one snapshot of its cases, and reject keyed each expected error by line number in one lib.rs, with every root appended to one list and a census whose totals change with any case. Two PRs that each add a case appended to the same lines, so they conflicted: of 71 merge-queue conflict ejections in four days, 43 were in lang and 28 in reject, and gate failure #374 was reject's line numbers shifting within a batch. A targeted lang run also translated all of lang.

Approach

• fixtures/lang/ is now a fixture per module: 66 crates, each with its own cases, ts.toml, expected/, and caller.ts when its cases name inputs. fmt, generics, and shapes share types, so they stay together in shapes. They share harness.rs (the case picker), lang.ts (the caller and its input helpers), the translation build dir, and the wasm target dir. fixtures.rs declares one test per fixture, so just ts test lang::ints translates and runs only ints.
• differential.ts takes a caller's ops, the bytes that pick each case, and scales the random-input count by them, so each case still gets 151 random inputs. The full-width u64 inputs it used to make for the first 8 ops, which were ints' ops, move to ints/caller.ts as boundaries().
• reject/src/ splits into a module per kind of construct. expected.txt keys each error by the item it points into and its offset from there (strs.rs:cow_matched+2:4), with closure spans keyed the same way, and keeps them sorted. The roots in ts.toml are module-qualified, and a .taplo.toml rule keeps them sorted (just check fails on unsorted roots, just fix sorts them). bind and crates, whose refusals share rejected(), get the same keys.
• The census over reject is still tested. It lowers generic items that no root reaches, which finds refusals translation never reaches, such as in config::generic and rebound::generic_park_rebound. Its spans are now snapshotted in reject/census.txt, keyed and sorted like expected.txt. The census's totals and tables, which changed with every reject case, move to a new fixtures/census crate of a few refusals that should stay refusals, so they change only when the census report itself changes. sysroot uses that crate too, since it depends on bytes.

Demonstration: one script made two branches from each layout. Each branch adds one lang case (in ints on one branch, strs on the other) and one reject case (pointers on one, fmt on the other), with snapshots regenerated by the real driver.

Layout	Files changed per branch	Merging one branch into the other
main	8	conflicts in lang/src/harness.rs, lang/expected/harness.ts, reject/src/lib.rs, reject/ts.toml, reject/expected.txt, reject/census.md
this PR	8	merges cleanly, and the merged branch passes reject, reject_census, census and both lang snapshots unchanged

Translating one lang module in that script took 3-8s, against 3-4 min for all of lang on main.

Expected snapshot changes:

• Translating each module apart changes a few names in the generated TS. For example, an import alias counts_Noisy is now Noisy, and one copied() of Bumped in iters is elided now that the iters crate alone decides how Bumped is copied. Each module's behavior is still checked against native Rust.
• reject's messages name types by module path, for example pointers::Halves, because those types moved into modules. Apart from location and module prefixes, the errors match main's one for one (checked with a script).

Impact

• Public API: none. Fixtures and tests only.
• just ci on the final head: 131s wall, of which the fixtures test binary took 112s for all 78 tests. On main, the one lang test alone took 7.5 min wall and about 11 CPU-min. The 66 lang tests take about 9 CPU-min and run in parallel.
• Open PRs that touch these fixtures need their edits moved to the new files on rebase. Moved src/*.rs files follow by git's rename detection. Folds from the back, and a crate's own folds #463's edits to lang/caller.ts move to lang/ends/caller.ts, and its reject/src/lib.rs cases move into the matching module, with roots module-qualified. fix(ts): drop a string search's or retain's closure where std does #471, fix(ts): translate by_ref of a mutable reference to an iterator held in a local #475, and feat(ts): translate Arc and Rc of dyn Iterator, and dyn FusedIterator #478 have landed and are merged in here: fix(ts): drop a string search's or retain's closure where std does #471's removed reject cases leave strs.rs, fix(ts): translate by_ref of a mutable reference to an iterator held in a local #475's iters::lent case moves to the iters fixture, and feat(ts): translate Arc and Rc of dyn Iterator, and dyn FusedIterator #478's dyns cases regenerate in the dyns fixture.

Alternatives

• Keep one lang crate and split only its case list and caller into per-module files: this removes the list conflicts, but each case still changes the one shared snapshot of its cases, and a targeted run still translates everything.
• Snapshot the census over reject as before with item keys: its totals and per-construct counts still change with every case, so any two reject PRs would still conflict. Dropping it would lose coverage of generic lowering.
• Sorting reject's roots by hand instead of through taplo: nothing would enforce it, and appends at the end would come back.

Follow-ups

• The go and ffi reject fixtures key errors by line number the same way. No merge conflicts have been counted there, so no quest was planned.

Review

Round 2: Codex, gpt-6.1-sol, medium effort, quest review --base origin/main on head ba6d201091d01ba0361a1b994e8136ca580b1c05: no findings. Round 1 on cb659c55 also had none.

🤖 Generated with Claude Code