[quinnj]

Builds on #1222; merge that first. This branch contains #1222's commit: the copy below widens the window in which a chunk parses under a stale type, and without #1222 a typemap target outside the promotion order then turns a sparse column into text on every parallel read (an independent review measured 40 of 40 runs; main already does it in about 30 of 40).

directchunk! took the shared promotion lock once per column per chunk to read that column's current type and final vector. A 1,000-column file in 118 chunks took it 118,000 times per read, and that wait showed up in 8-task profiles. Each chunk now copies every column's type and final once, under one lock.

Why a stale copy is safe: if another chunk promotes a column after the copy, this chunk parses that column under the old type into the old final, records the old type in segtypes, and directwave! re-parses it into the promoted final after the wave (redirect!; a chunk that parsed as Missing only refills its slice). That is the existing path for a promotion that lands while a chunk is parsing the column; the copy widens the window in which it can happen, which can only add re-parses, and only when a promotion races with the copy. A chunk's own promotions still read the current type under the lock. The cost: a chunk whose copy predates a promotion parses that column under the old type before the re-parse and keeps the replaced final vector alive until the chunk finishes; both happen only when a promotion lands during the wave, which sampling makes rare.

Interleaved process pairs, Julia 1.12.6, M3 Max (load average about 8):

file	tasks	main	this PR
wide_numeric (124 MB, 18k rows x 1,000 columns)	8	0.057 / 0.085 / 0.071 s	0.045 / 0.063 / 0.057 s
same, total CPU	8	0.40 / 0.52 / 0.53 s	0.30 / 0.39 / 0.38 s
wide_numeric	1	0.38 s	0.38 s
long_numeric (5M x 8), mixed (2.5M x 10)	8	within noise	within noise

Review: my own pass, then an independent agent review (stale detection, the free Missing upgrade, String finals and adopted slots, limit, Timestamp widening; 144 parallel runs matching serial), which found the typemap interaction now fixed by #1222.

• Traced every consumer of the per-chunk state: stale detection compares segtypes[k][j] with the frozen final type, so any chunk that parsed under an old type is re-parsed; String finals adopt each chunk's bytes into the chunk's own slot after the wave; Timestamp nanosecond-to-microsecond widening reads only segtypes and the source spans.
• The snapshot is taken after the chunk's structural row checks and before any column parse; nothing between them reads promo or finals.

Verification:

• New testset promotions that land while another chunk parses: nsample=1 seeds every column from the first row, so all 40 columns promote across 948 chunks; four parallel reads must equal the serial read in values, types, and problems, with and without a typemap. It guards the re-parse path this change leans on (the speedup itself is measured above, not tested).
• Full suite --check-bounds=yes -t4: 331,812 passed (head, including Re-parse a stale chunk under the column's current type #1222).
• 300 seeds of a cross-configuration consistency fuzzer (every reader, task count, and chunk size), 150 seeds of a feature fuzzer, and a promotion probe (one odd value at the first, middle, and last row of 200k, at 1 and 8 tasks): no differences.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 95.79%. Comparing base (2f84f42) to head (44886f5).

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #1221   +/-   ##
=======================================
  Coverage   95.79%   95.79%           
=======================================
  Files           9        9           
  Lines        5566     5570    +4     
=======================================
+ Hits         5332     5336    +4     
  Misses        234      234           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.