[morsel-ray]

Landed normally

02804c3: independent Anton approval of the prefetch feature preserved through normal reviewed-base imports; fresh currentf7/head/APPROVED/resolved/CLEAN/MERGEABLE and all11 actual COMPLETED/SUCCESS gates before ordinary protected expected-SHA squash. Complete checked/landed tree f2f264a91d1fb7c8bdddbf5b8e931f1a690d3eb0 identical; remote dev verified. No master/tag/release action. No pool-policy/index-routing issue auto-closed (closing references checked empty).

Combined-head local join114/identity3at0/4helpers and other scopes below retain their actual provenance; earlier prefetch-only native measurements are not remeasured whole-combined-head prices.

Current accepted-identity-LEFT integration — 41abae79

Normal expected-head import of accepted #943 /f7b23ec7. Ordered parents/computed complete merge tree verified. join.c combines accepted #943 output-retention and per-column slice fallback; all5other feature blobs unchanged. Complete incoming join.c delta personally read; the prefetch feature is preserved, not a byte-identical whole join.c. Fresh non-root ALL-TU ClangASan/UBSan join114, no-helper pair1, affine20, zeroRFL2, imported identity3each at0/4configured helpers, TIME3/metadata2 passed. GCC DEBUG/NDEBUG syntax2, whole relevant logs diagnostic0, exact clean source/both binaries checked. Compared with accepted #943s113, the sole additional selected case is rfl/join/tiny_chained_prefetch. No new local full/TSan/combined-head native/backtester claim. Fresh actual exact-head/current-base/review/discussion/gates remain required before landing.

Prior TIME32 integration — cee4db5f

Normal expected-head import of accepted #939 /5f6e619b. Ordered parents/computed whole merge tree verified; all6feature blobs unchanged. Fresh non-root ALL-TU/header-consistent ClangASan/UBSan join110, no-helper pair1, affine20, zero-divisor2, imported TIME3 and metadata2 passed. GCC changed production TU DEBUG/NDEBUG syntax2, whole relevant logs diagnostic0, exact clean head/both binaries checked. No new local full/TSan/native/backtester claim; earlier evidence retains its named heads. Fresh exact-head CI/current-base/reviews/threads remain required before ordinary landing.

Prior B0 integration — bf022ae7

Normal expected-head import of accepted #936 / b0b0a1b. Ordered parents and computed complete merge tree verified; all6feature blobs unchanged. Fresh non-root ALL-TU/header-consistent ClangASan/UBSan join110, no-helper pair1, affine20, zero-divisor2 and imported metadata2 passed; GCC changed production TU DEBUG/NDEBUG syntax2, whole relevant logs diagnostic0, exact clean head/both binaries verified. No new local full, TSan, native pricing or author-backtester claim. Earlier measurements/controls below retain their original heads. Fresh exact-head CI is still required.

Prior D715 integration — a1a4bcb7

Anton independently approved 5bbe455. Normal import of accepted d7153e7 / #937 givesa1a4bcb7: ordered parents/computed complete merge tree verified, all6feature blobs unchanged, including join.c/fixture/benchmark files. Fresh non-root all-TU/header-consistent ClangASan/UBSan join110/110, no-helper pair1/1, affine20/20, zero-divisor2/2, imported typed-sort probe1/1; GCC join DEBUG/NDEBUG syntax2. Whole relevant logs diagnostic0, exact clean head/both binaries checked. This does not relabel earlier full/Tsan/native measurements as new executions. Waiting for current head’s actual required checks; no self-approval or bypass.

Why

The chained join's count and fill passes hash row i+8 again merely to prefetch a bucket head. With the minimum256-head array (1KiB), that speculative work is expensive on the small-build workload measured here.

Change

• Name the existing minimum capacity; choose count/fill prefetch-mode callbacks once per pass.
• Shared inline bodies receive constant mode arguments. Minimum-capacity callbacks omit only future-row hashing/head prefetch; actual current-row hashes, equality, chain walking and output stay unchanged.
• Both serial/pool paths use the selected callbacks. Larger-table and build-side prefetch, allocation, task geometry, existing domain warmup, cancellation and ASOF/radix routing are unchanged. No new type whitelist, cache or validation pass.

The gate uses capacity, not distinct-key count: right128/129 selects256/512heads. A small head array does not prove the entire working set resident.

Measured, bounded release result

Matched Clang18O3/native/i7-6700 comparison: immutable acceptedbda2f218 release, replace only its join object with the exact candidate source; common objects and flags/hash proofs retained. This is not a fresh whole-final-head release binary. The pre-#943 candidate matched the measured TU. Current combined join.c additionally contains accepted #943; earlier prices are not whole-combined-head measurements.

Fixed130k-row/773-symbol composite-key input,200 inner joins against3unique keys, independent count101400/key checks; selected-core affinity, no concurrent local builds/tests, randomized fresh processes/eight samples after separate discarded warmups. Requested participants are not instrumented worker participation.

Total participants	Process wall, baseline→candidate	UserCPU	Join-loop ms
1	1.666→0.844s	1.662→0.840s	1654.505→833.265
4	0.488→0.261s	1.819→0.933s	476.759→249.971

A three-way162-process comparison also priced the simpler row-mode-test version: callback selection improves its one-participant loop854.914→833.265ms; four-participant255.175→249.971ranges overlap. Inspected Clang objects have the mode-specific prefetch eliminated/retained as intended; object text grows1240bytes. Not a universal compiler guarantee.

Retained-prefetch controls are not hidden: SYM129 candidate loop282.951/88.069ms versus baseline281.815/87.970, longSTR1292151.520/691.559 versus2159.760/691.727 (1/4participants); ranges overlap, including a large serialSYMoutlier. The earlier simpler candidate's broader11-case study had a radix-size65537/four-participant+4.8%median with overlapping ranges; that is not final-version radix pricing. No universal neutrality, new instruction-count result, exclusive-host or whole-engine speed claim.

Correctness and provenance

• Exact pair-ID RFL checks duplicate/null/missing matches for INNER/LEFT/FULL, the128/129capacity boundary and8197probe rows across a morsel. Timing count checks do not replace pair equality.
• Historical private current-D9 headadb16ece, same runtime/fixture blobs: forced join110/110; full non-root ClangASan/UBSan 6419/6436passed,17skipped,0failed. All1488tracked-source manifests/both binaries/HEAD/index guards and independent whole-log diagnostic checks match; binaries preserved.
• Final5bbe4556 imports accepted5f44d2d1, then changes only benchmarkREADME. Computed merge tree and all six feature blobs before that docs-only edit match. Fresh consistent-header all-TU final-head Clang ASan/UBSan: join110/110, no-helper pair fixture1/1, imported affine20/20 and zero-divisor2/2. GCC join.c DEBUG/NDEBUG syntax2pass; forced current scopedTSan pair fixture1/1 with4background helpers configured (participation uninstrumented). Whole relevant logs diagnostic0, exact clean head/canonical identity/both normal binaries checked. This is not a second full-final-head run.
• Source-only lifetime review found no new required side effects in removed speculative hashing under existing warmed-SYM/STR/domain/materialization contracts; not independent maintainer approval or general cancellation certification.

bench/join_tiny_ht supplies a fixed-key probe and standard-library Python process-comparison wrapper; preserve separate release binaries (test builds overwrite the CLI). This is related to the workload discussed in#599, not a scheduler-policy fix or resolution of that issue. No spill/domain/cancellation contract expanded or allocator-failure injection claim. Independent maintainer review and fresh required CI remain necessary.