I want three physical representations behind the same world and session API. The string format remains useful, but a fixed-width key should not pay for front-coded string navigation. This is the format direction; .kv is the implemented persistent representation, while .fv and .ff are additions to build.

File	Keys	Values	Native offset directory
`.kv`	Front coded or sort-owned mixed grammar	Sort-owned grammar	Sampled Elias–Fano record offsets
`.fv`	Fixed-stride key array	Variable-size packed payloads	Elias–Fano value boundaries
`.ff`	Fixed-stride key array	Fixed-stride value array	Arithmetic addresses

For .fv, key ordinal \(i\) has an arithmetic address. A monotone sequence of \(N+1\) value boundaries selects the half-open payload interval for that ordinal, including the real end-of-file boundary. Repeated offsets can represent empty payloads; the sort must still distinguish empty values from absence. The Elias–Fano universe includes the value payload extent, not the fixed key array. This directory samples every value, so its select latency and storage tradeoff differ from the current every- \(W\) record directory.

The offset-directory design makes the representation a per-file choice: Elias–Fano, packed absolute positions or direct positions. The sort registry still owns record semantics. That format choice remains a design proposal; current persistent files use Elias–Fano.

The fixed-key Metal experiment constructs complete native outputs for candidate .fv and .ff layouts. Its fixture envelope is experimental; these formats are not yet persistent runtime choices.

For .ff, both addresses are arithmetic and there is no native Elias–Fano directory. Each fixed-stride region needs one physical slot width. How mixed sorts share those regions belongs to the wire-format design; per-sort widths alone do not establish one common stride.

Fractional indexes

The fixed-key formats still sample the downstream augmented native-plus-index catalog. Their fractional indexes hold fixed-width borrowed keys, grouped rank, and equality/false-borrow information. They need no index-side Elias–Fano directory and no inherited FC comparison context. Once rank identifies the native or borrowed ordinal, its key address is arithmetic. COLA main and terminal secondary routes retain their existing meanings.

The .kv representation retains its sampled native offset directory and a separate directory for each front-coded borrowed stream. Changing the physical key representation does not change what the downstream sampling refers to.

Searching fixed-key windows

<everett/fixed_search.h> supplies a bounded search over the flat key arrays. fixed_key_view<1>, <2> and <4> read 32-, 64- and 128-bit keys represented as one, two or four little-endian unsigned 32-bit words. Comparison follows word order, with word zero first, as in the fixed-key merge experiment. A raw native-endian 128-bit integer has a different ordering.

everett::fixed_key_view<4> keys(mapped_key_bytes);
auto window = keys.subview(first, count);
auto local = window.lower_bound(query);
if (local != window.size() && window.key_at(local) == query) {
  auto ordinal = first + local;
  // Locate the value for ordinal in its separate payload region.
}

The view borrows its exact byte span, allocates nothing and supports unaligned starts. upper_bound handles the other equality boundary. Explicit _binary and _simd variants make the choice measurable. SIMD searches subdivide small windows with parallel pivots, then compare a bounded leaf; larger windows use binary search. These are key-array primitives, not yet a complete fixed-format fractional-index reader.

AVX2 and AVX-512 use fault-suppressing masked loads for partial vectors. NEON and SSE2 load only the valid bytes in a partial vector. Crossing a readable page is allowed; reading beyond the supplied span is not. The tests put spans against protected pages, including empty spans and short final windows.

I keep 15:1 sampling independent of this search choice. A virtual sampling cut projects through rank to an arbitrary native or borrowed ordinal, so aligning the beginning of a key array does not align every projected search window. Khuong and Morin's array-layout study motivates comparing flat branch-free search with alternatives; its machine measurements do not establish the best cutoff for these short windows. The bounded search measurements select conservative NEON and AVX2 cases. Other windows use scalar binary search by default; the explicit variants remain available for measuring a different host or workload.

Logical keys and shared machinery

Fixed keys can be structured IDs, preserving useful entity/component or region order, or hashes of stable logical identities. Content addressing is a separate choice: a stable logical key can refer to a changing content hash. The weak world signature is not a substitute for an identity or collision policy.

Sort dispatch, logical ordering, composition, signatures, pins, snapshots, sessions, durability and merge scheduling stay shared. Physical access and merge kernels specialize for the table representation. Fixed-width keys permit direct parallel input cuts; variable output payload sizes still require sizing, prefix sums and copying. Tombstone elision still requires complete older-history coverage, regardless of key encoding.

The current Elias–Fano implementation is tuned for sampled offsets. Before choosing its .fv layout, I want to measure finer select directories and paired adjacent-boundary access for per-value navigation. Sequential merge traversal also needs its own baseline: repeatedly selecting independent offsets can cost more than walking the encoded sequence once.