chore(B7): defer plan; tuple cost crossover makes large tables a regression

[davydog187]

Defer B7: tuple cost crossover makes large tables a regression

Updates .agents/plans/B7-table-array-hash-split.md from ready → deferred with the measurement data that drove the decision to close PR #229 without merging.

Why we closed #229

#229 implemented the full plan correctly (all 1692 tests + 29 lua53 suite tests passed). But when we re-measured under the bench harness from #230 with multi-n inputs, the picture changed:

workload @ n	main	B7 (#229)	delta
Build n=100	17.09 µs	14.03 µs	-18%
Build n=1000	197.96 µs	265.82 µs	+34% ⚠️
Sort n=100	34.91 µs	27.57 µs	-21%
Sort n=1000	490.49 µs	655.72 µs	+34% ⚠️
Iterate n=100	24.59 µs	21.11 µs	-14%
Iterate n=1000	276.74 µs	358.64 µs	+30% ⚠️
Map+Red n=100	49.79 µs	42.78 µs	-14%
Map+Red n=1000	603.93 µs	843.57 µs	+40% ⚠️

Memory regressed 3-5× at n=1000 (Sort: 2.08 MB → 12.40 MB).

The crossover is structural: BEAM tuples are immutable, so every setelement/3 on a 1024-cell tuple copies the whole tuple. PUC-Lua avoids this with in-place C mutation; we can't. Small n wins because setelement/3's constant-factor advantage over Map.put matters; large n loses because copying dominates.

The single n=500 measurement that originally motivated B7 was right at the crossover. Multi-n in #230 was what made the pattern visible.

Conditions for reconsidering

A future plan could revisit this with threshold-based promotion: keep contiguous integer keys in the hash map until array_len reaches some boundary (e.g. 256), then promote. That would preserve the small-table wins (-14% to -21%) without taking the large-table hit. The plan file documents what such a plan would need to address.

Until that plan exists and ships, table workloads have to look elsewhere for durable wins — likely B5 (compile prototypes to Erlang) which attacks setelement/3 register-write cost directly.

Changes

 .agents/plans/B7-table-array-hash-split.md | 73 +++++++++++++++++++++++++++--
 1 file changed, 70 insertions(+), 3 deletions(-)

Only the plan file changes; no library code touched.

Verification

mix format
mix compile --warnings-as-errors
mix test

(All trivially pass; this PR only edits a plan file.)