perf: detoast each varlena once per row on the write path (#445)

[jdatcmd]

What

PgColumnarWriteRow flattened every varlena value up to four times per row — once in PgColumnarEncodeValue, once for the bloom hash, and once per min/max comparison. For a toasted (pglz-compressed) value each of those is a full decompression; #445's write-path profile saw it as detoast_attr.

This detoasts once at the top of the row's column handling and reuses the flat Datum for the encoder, bloom hash, and min/max compares, freeing the copy after the row. pg_detoast_datum returns the same pointer for an already-flat value, so the uncompressed common case copies and frees nothing and non-varlena types skip it — output is byte-identical either way.

Measured

300,000-row load of a ~3.8 kB pglz-compressed text column, pg18 non-assert, median of 3:

	load
before	3193 ms
after	2836 ms

11.2% faster. The saving scales with how compressed the values are — small for short strings (ClickBench URLs), larger for wide text. This is one item in #445's broad write-path tail, not its headline (that is compression, which parallelises).

Tests

test/native_toasted_write.sh (new): loads a heap table's toasted values into columnar and differences against a heap mirror — values byte-identical, min/max range prune correct, equality (bloom + recheck) correct, and the short non-toasted column unaffected. It guards the flatten-once path specifically, which the other write suites don't reach (they use short, non-toasted values). Registered, harness_selftest green (107).

Correctness unaffected: differential, native_writer, native_zonemap, native_bloom all pass. The removal proof for the optimisation is the load-time number above (deleting it stays green — the output is identical — so the perf win is a timing, not a check).

🤖 Generated with Claude Code

[ChronicallyJD]

Approve — verified on four dimensions, prove-not-trust

Reviewed independently of the author. The detoast-once change is correct, memory-safe, faster, and adequately guarded.

• Correctness vs heap oracle. differential + native_toasted_write pass. I added an external-TOAST probe the base fixture doesn't reach — 4160-byte values with STORAGE EXTERNAL, premise-verified is_external=t — and columnar matches heap on values, min/max pruning, and bloom equality. That closes the one nuance in the change: datumCopy on the zone-map min/max now copies the flattened value rather than a possibly-external pointer, and it prunes identically to heap on the external branch.
• Memory-safety. Forced ASAN+UBSAN (instrumented .so confirmed): native_toasted_write, differential, native_writer, native_dml, write_fsst_compressed, encode_invariants, corruption, native_reclaim — 0 sanitizer reports. The new per-row pfree(flat) is clean.
• The guard is non-vacuous. Removal proof: corrupting the flatten (marker verified live in the .so, not stale) flips native_toasted_write to FAILED on the value/bloom/non-toasted checks.
• Perf reproduces. Interleaved A/B, 300k-row toasted-text load, pg18n non-assert, 6 reps/arm: ~8.3% faster, every rep separated (your 11.2% on a more concentrated wide-text column; mine is diluted by a 3-column table). Direction unambiguous.

Merge remains yours — this is a review approval, not a merge.