[sjkim1127]

[AUTOMATED]

Scope

Mechanically accelerate cold FID/no-return preparation and immutable runtime metadata lookup. No semantic/readability/score/ranking claim, CLI/schema/adapter/dependency change, skipped function, modified instruction limit or FID matching rule.

• Borrow an indexed canonical registry descriptor for hot status/engine queries and hoist invariant endianness; preserve the public owned API, duplicate manifest order, exact original selection predicate and missing-entry errors.
• Only seed-capturing cold FID preparation uses bounded scoped OS helpers; ordinary no-return callers remain sequential. Immutable per-function decoding cannot query CF/no-return. A local atomic queue assigns original positions once; the caller works and joins without stealing outer Rayon jobs under the CF OnceLock.
• Existing Rayon width includes the caller; process-wide nonblocking permits bound extra helpers. Thread creation/permit failure reduces parallelism without dropping work. Restore original input order before map insertion. Preserve fixed point, raw seed/complete decode cache contracts and fallback behavior. RAII returns permits and propagates panics.

Measured evidence

Base main 3daf458fef337a2951e000a17f13fac61d4a2e24; frozen qualification revision 07dd95dbe6214b8a78fc1e4a55961ab515986489.

Large six-function native batch, three alternating baseline/candidate pairs, existing default rendering/FID workers; medians:

	Baseline	Candidate
Wall seconds	27.2216	15.4257
CPU seconds	122.0768	136.2194
Maximum RSS bytes	2,329,264,128	2,416,246,784

Wall -43.3%, CPU +11.6%, RSS +3.7%. This lowers latency at higher CPU cost. Small five-pair median changes are only 1.6–6.7%. No per-core efficiency or memory-cap claim. Native M4 Pro/24 GiB, release LTO; OS caches/scheduling uncontrolled. No own builds/tests/profilers during primary timing controls.

Fresh 250/250 stripped functions across 45 cells: zero changed non-time function records, including all code/layers, variables/types/PCs/maps, errors/fallbacks/evidence/counters; only seven existing durations excluded. Focused 22-function complete records also match three repetitions in each of two worker lanes and the fresh replay. Whole-population two-run determinism, guest execution/recompilation and rescoring were not measured.

Kuna 1.808 is compared with both 1 and 14 requested workers. Choosing each tool's faster observed setting separately still leaves Fission 2.83–7.24x slower; Kuna-level speed is not achieved. One requested Fission render worker does not serialize FID/preparation. We do not pick Kuna's slower 14-job setting to claim parity.

Validation

• 17 focused nextest tests; 1,855 relevant sleigh/static/pcode/decompiler/CLI tests passed, 6 skipped.
• Registry differential: every manifest entry, invalid/duplicate keys, borrowed identity and owned independence.
• Queue ordering/every-input-once/skips, caller-only fallback, real OS spawn failure, unavailable permits, panic recovery, concurrent helper cap; saturated outer Rayon/OnceLock stress (24 rounds × 24 callers, explicit deadline) completes with exactly one initializer each round.
• Serial/scoped no-return facts and all 16 captured fixture decodes match exactly. Existing raw context/bytes/delay/partial-capture coverage remains green.
• Crate checks, release CLI, fmt, owner boundaries and scoped CI Clippy passed.
• Separate prototype A1 symbolized samples (before the final CAS portability fix) expose remaining FPK lookup/decompression/candidate parsing; these are diagnostic snapshots, not timing or whole-run attribution. Source hashes/patch/tree, full ranges, controls and reproduction: docs/validation/preparation-workers-performance.{md,json}; proposal docs/proposals/2026-10-15-preparation-workers.md.
• Full artifacts: benchmark/artifacts/performance/2026-10-15-preparation-workers/.

All-function hot-cache growth remains unchanged. Drafts #203/#204, DecBench #103, primary checkout and seven unrelated dirty main files are preserved. Remote CI is a separate post-submission gate and is not claimed green by local tests.

CI portability follow-up

Initial CI passed Linux tests and PE CLI smoke; lint failed because newer CI Rust deprecates fetch_update. An equivalent explicit compare_exchange_weak loop preserves the helper limit and memory order without a new API requirement or warning suppression. All timing controls, 250-function replay and both Kuna worker lanes were rerun on final A2 SHA256 5fde8204f59b90a4fba7ca2af4c32c08c41efcb066df8586a9085cf7022f65d7. The table above now reports A2; older prototype controls/samples remain explicitly separate. New-head remote CI is still required.