refactor(code-index): route chunk fan-outs through one yielding helper

[ScriptedAlchemy]

Summary

• Move both chunk fan-outs in chunks.rs (try_for_each_chunk_ordered, admit_all) behind one private fan_out::map_yielding that performs the with_yielded_background_cpu_permits wrap internally.
• rayon is now imported only inside mod fan_out, so a bare par_iter anywhere else in chunks.rs fails to compile.
• Add a source-scan test that pins every pool token in chunks.rs to mod fan_out and pins the yield to that module.
• Isolated tip leaf forked once from 9fbf169fe; not re-rebased as the tip moves.

Motivation

The nested-admission deadlock fix from #1449 (f81df3b34) was two call-site wraps around two par_iters. The merge resolution in d039bee89 (#1435) kept both par_iters and dropped both wraps; the base stayed wedge-prone for sixteen merges (#1435 through #1383, including 630702adb where the hang was re-observed) until 76a8e7838 (#1460) restored them. A fix that a conflict resolution can silently detach from the code it protects is not durable. Welding the yield to the fan-out inside one helper, and making that helper the only route to the pool in the module, means the yield cannot be kept out while the fan-out stays in.

Changes

• crates/tracedecay-code-index/src/chunks.rs
  • New private mod fan_out with pub(super) fn map_yielding<I, R>(items, map) -> Vec<R> for any IntoParallelIterator whose iterator is indexed. It yields the caller's admitted units around the join and collects outputs in input order.
  • try_for_each_chunk_ordered: map_yielding(chunks, |chunk| admit(&mut || operation(chunk)).err()) then the first Some in input order. This equals the previous filter_map + min_by_key(index) because indexed collect preserves order.
  • admit_all: map_yielding(chunks, |chunk| with_background_cpu_permit(|| self.admit(chunk))).into_iter().collect().
  • Top-level use rayon::prelude::* removed; the import lives inside mod fan_out.
  • Test every_pool_fan_out_in_chunks_goes_through_the_yielding_helper: include_str!("chunks.rs"), skip comment lines, assert no rayon / par_iter( (also matches into_par_iter() / par_chunks / par_bridge token outside mod fan_out, and assert the module contains with_yielded_background_cpu_permits(. Token literals are built with concat! so the test's own source is not an occurrence. It caught itself on the first run (identifiers named rayon_tokens), which is the evidence it scans what it claims to.
• Sequential path (< PARALLEL_CHUNK_THRESHOLD) untouched. No behaviour change on the parallel path beyond collecting Option<ChunkingFailureV1> per chunk instead of filter_map.

Test plan

All on the fork point 9fbf169fe, --profile perf, CARGO_INCREMENTAL=0, own target.

Check	Result
cargo test -p tracedecay-code-index --lib whole binary ×3	246/246 ok ×3 (7.8 s, 7.0 s, 7.5 s); nested-admission pin and the new guard green in each; one more 246/246 after commit split
Guard self-check	first run FAILED on the test's own rayon_tokens identifiers before the rename; passes after — the scan sees real code lines
cargo fmt --check	no diffs in chunks.rs
cargo clippy -p tracedecay-code-index --profile perf --no-deps --all-targets --keep-going -- -D warnings	lib, lib unittests and every other target clean; the only diagnostics are two pre-existing E0425 in tests/code_index_suite/chunk_incremental.rs:362,366 (base red, reproduced identically with this diff stashed)
npm run lint:commit -- --from origin/codex/... --to HEAD	pass
cargo test -p tracedecay-code-index-runtime --lib -- reconcile	not runnable on this tip: the runtime crate does not compile at 9fbf169fe (publication_store.rs:2640 .reused), identically with this diff stashed; being fixed separately. The same filter completed 2/2 (168 tests) on d1586e4ae earlier today with identical wrappers in place.

• cargo nextest run --workspace --no-fail-fast — CI (expected red until the base compile fixes land; merge only on real SUCCESS)

Checklist

• CHANGELOG.md — internal refactor + test, no user-facing change
• No secrets, credentials, or .env files included
• No breaking changes

Base reds seen while verifying (not fixed here)

1. crates/tracedecay-code-index-runtime/src/code_index_scheduler/publication_store.rs:2640 — .reused on ChangedCodeChunkSetV1; the runtime crate does not build at 9fbf169fe.
2. crates/tracedecay-code-index/tests/code_index_suite/chunk_incremental.rs:362,366 — parallelism::force_install_failure_for_test is no longer exported from the lib to integration tests.
3. code_index_scheduler::tests::reconcile::concurrent_query_admissions_claim_one_pending_wake_before_worker_coalescing fails alone 2/3 on d1586e4ae (admitted == 0); present since d62bf238d, unrelated to background-CPU admission.

  

[changeset-bot]

⚠️ No Changeset found

Latest commit: e88be5e

Merging this PR will not cause a version bump for any packages. If these changes should not result in a new version, you're good to go. If these changes should result in a version bump, you need to add a changeset.

Click here to learn what changesets are, and how to add one.

Click here if you're a maintainer who wants to add a changeset to this PR