[VibeSWE]

For fill_null=true with nonempty, unique requested fields, append missing nulls by direct requested-field index. Reuse the existing map and preserve the original writer body for duplicate or empty requests and fill_null=false. This avoids repeated map lookups in the selected null-fill loop. Add six regression tests covering byte order, missing/duplicate fields, nullable rows, changing schemas, empty/columnless batches and fallback behavior.

Validation uses retained evidence for the exact production and test sources: 22 baseline + 22 candidate tests passed in the component harness; nine fixtures showed exact JSON and heap parity over 20 calls per arm. The component harness uses a DataFusion shim and a separate compilation profile. The ordinary-release combined application completed the full campaign: 18 lifetimes, 90 cells and 84,600 timed responses; all native exits were zero and closure was independently verified.

In that campaign, C/B isolates this serializer on the same correctness-fixed source and matched plans/bytes/queries. Across three blocks, process CPU/query fell 4.9–5.2% for H-fill and 10.6–11.3% for W-unique. Default-path performance remains unresolved (C/B 0.997–1.011); application memory remains unresolved. These are descriptive results from one host with fixed datasets and concurrency 1, not a blanket no-regression or adoption claim. This PR contains only the serializer and tests; it is separately reviewable from the sorting correctness fix.

Full-crate validation now passes on 026c0615d896b297ee9e9cda80acecfae3d25daa: Rust tests, Clippy, image build, and standalone/distributed smoke tests are green (CI run). The standalone branch has not been separately performance-measured; the measured application was the combined build including the reader fix in #48. The retained component measurements and the full-crate CI checks are distinct evidence.

The helper deliberately copies the original writer body to keep the fallback source unchanged; future writer fixes must keep the two paths consistent. The user authorized merge on 2026-10-21 after reviewing the disclosed tradeoff. PR #48 is merged. CI passed again on the combined head 6acbbffe2f6a04e2fb2a02344902f5493bb7d7ee before this PR was merged (combined CI run). Successful CI does not establish application memory parity.

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}