Add exact keyed collection reduction

[PraneethMerugu]

Outcome

Adds one exact fixed-capacity keyed collection reduction to LocalMath's existing StageProgram and KernelAbstractions executor. This is the reusable storage/publication primitive required by CorePotts maintained generation-aware owner-pair multiplicity without an O(C²) directory or collision-unsafe hash.

Contract

• exact Int32, UInt32, or bounded flat-tuple keys
• unique stage-entry keys
• existing value first, then participating contributions in canonical (source, lane) order
• explicit new-key identity and retain/drop-identity policy
• private O(capacity + source_count × maximum) workspace
• failure-atomic records and logical-count publication on invalid prior count, duplicate prior keys, invalid control, or overflow
• one shared CPU/Metal semantic implementation; no Potts vocabulary, registry, hash mode, relaxed fold, or second executor

Compiler and allocation evidence

• matched Collect host orchestration: 112 typed statements, 79 calls, 22 Any SSA slots
• keyed host orchestration: 222 typed statements, 165 calls, 31 Any SSA slots
• device-reachable materialization boundary: 129 statements, 73 calls, zero Any-valued expressions (37 control-flow/return CodeInfo slots)
• device-reachable sort boundary: 260 statements, 139 calls, zero Any-valued expressions (87 control-flow/return CodeInfo slots)
• device-reachable fold boundary: 363 statements, 204 calls, zero Any-valued expressions (105 control-flow/return CodeInfo slots)
• device-reachable publication boundary: 231 statements, 137 calls, zero Any-valued expressions (63 control-flow/return CodeInfo slots)
• capacity specialization count: 1 across capacities 4 and 8
• across four operation/retention variants, bounds/emission/sort/segment/finalize/publication each have one specialization; folding has four, as intended
• warm public call: 8,432 bytes versus 4,768 bytes for the matched Collect control
• the 3,664-byte delta is shared KernelAbstractions kernel/NDRange/keyword/argument construction and receipt bookkeeping; device workspace is preparation-owned and execution performs no device allocation
• the reproducible probes were executed through a Kaimon persistent LocalMath project session

Eliminating that host delta safely requires an executor-wide prepared-launch/fusion contract. This PR does not add a keyed-only shortcut or fuse across required global scan barriers.

Validation

• focused keyed CPU contracts: 40/40
• real Metal with scalar indexing disabled: 24/24
• full LocalMath package suite: 1,853/1,853
• documentation build, doctests, link checks, and rendering: passed
• public constructor/API regression after final tightening: 27/27
• independent review and final narrow re-review: clean

Chain position

Planning companion C08, based on merged LocalMath PR18 (9d3e1a24). It precedes Core R10 maintained spatial queries and Potts R11 lowering.