CultLib/tests/vectors/unique-index

unique-index vectors

Written only by the C# reference, tests/GameCult.Caching.Tests/UniqueIndexVectorTests.cs, with CULTCACHE_WRITE_VECTORS=1. Every other run reads these files and compares them byte for byte; a changed vector is a reviewed commit. Other runtimes are held to the same canonical text and the same verdicts (variants:ruling:unique-indexes).

Files

• values.msgpack: an array of [label, kind, value, text-or-nil], kind one of string, int64, uint64, double, bool (decode integers 64-bit wide). text is the canonical index text of a member holding value: a record holding it is found by GetByIndex(alias, text). A nil text is a blank value: no value, never indexed, never judged, found by no text, and two records holding it commit. Strings are as is (never trimmed), blank being empty or only C# White_Space characters (U+0009-000D, U+0020, U+0085, U+00A0, U+1680, U+2000-200A, U+2028, U+2029, U+202F, U+205F, U+3000); U+FEFF and U+001F are values. Booleans are True and False. Integers are their decimal digits. A non-integral finite double is its exact decimal expansion (written by C# RenderExactDouble, never hand-typed), -0.0 is 0, NaN and infinities are nil.
• duplicate-store.cc: a single-file store, schema tests.unique_item v1, slots [Sku: string, Category: string], stored-at pinned to 2026-01-01T00:00:00Z. Records item:a and item:b both hold Sku S-1; item:c holds Sku " " and item:d holds Sku U+0085. Written through a role Sku.
• blank-store.cc: the same, only item:c and item:d.
• pair-store.cc, surrogate-store.cc, types-store.cc, repull-before.cc, repull-after.cc, alias-store.cc, key-case-store.cc, key-punctuation-store.cc, surrogate-groups-store.cc, holder-case-store.cc: single-file stores written like duplicate-store.cc, through roles, for the rows of refusals.msgpack that open or pull them. Schemas tests.unique_item, tests.uq_pair (Sku and Alt both unique), tests.uq_alias (Sku and alt both unique), tests.uq_alpha and tests.uq_beta, each slots [0: Sku string, 1: second string], schema version the name plus .v1. The schema ids C# writes in the store catalogs and on every record:
  • tests.unique_item: sha256:93bb890de83069061af696b23f94fd32fb3c2dad3c91ee645e8add56085f1882
  • tests.uq_pair: sha256:2538cceafb6dcbce22ef11b932ecb804af5f96bc3aaedff692a52d3640433e60
  • tests.uq_alias: sha256:77a77a394ebd8bf71e14b9b6f2cea4978c25d4e705fd12364f660edbf6c63b1f
  • tests.uq_alpha: sha256:53aeaf2ea9b25e85ade6b15fc04c7653e7b206c87e258f1aa74dd32a840a1389
  • tests.uq_beta: sha256:45ed62a64c91f6d815d3ec975c4a4a1b2ce07b6e7262a32d828673cf10899586
• outcomes.msgpack: rows [file, unique alias or nil, open|refused, k1 or nil, k2 or nil]. Opening file with that alias a unique index (nil: a role, no unique alias) yields the verdict. A refusal reads, byte for byte, Records <k1> and <k2> (tests.unique_item) both hold unique index '<alias>'; an index value names one record. with k1 < k2 ordinal, never contains the value, and leaves the file byte-identical.

refusals.msgpack

RULE: a unique-index refusal is a function of the state the admission would leave, never of staging, file or pull order. Over every (type, unique alias, non-blank value) that state holds at two or more distinct keys, name the one least by (alias, its least key, its second-least key, schema name), each compared by UTF-16 code unit (C# StringComparer.Ordinal): Records <k1> and <k2> (<schema>) both hold unique index '<alias>'; an index value names one record. with k1 < k2 its two least keys. One rule for a single write, a batch, a load, a re-pull and an index registration.

Rows [label, ops, refusal]. Each op is ["open", file-or-nil] (copy that vector file to a fresh path, or nil for a fresh path with no file, and open a cache on it), ["pull", file] (overwrite the open cache's file with that vector file, then pull every backing store again) or ["commit", [[key, schema, sku, second], ...]] (one batch, staged in row order; schema is the schema name). refusal is nil (every op succeeds) or [k1, k2, schema, alias]: every op but the last succeeds and the last refuses with exactly the RULE text built from those fields, leaving the file byte-identical. Schemas are those of the store files above; second is Category (plain) except in tests.uq_pair and tests.uq_alias, where it is Alt (unique, alias Alt and alt respectively).

The 13 rows of the spec are a floor; rows 14-24 extend it with orderings that a code-point, case-insensitive or culture compare would flip. Rows, in file order: 1 batch-later-value-staged-first, 2 batch-surrogate-keys, 3 batch-interleaved-types, 4 batch-two-aliases, 5 held-then-landing, 6 held-in-the-least-group, 7 swap-off-a-held-value, 8 load-two-aliases, 9 load-surrogate-keys, 10 load-interleaved-types, 11 repull-held-pair, 12 load-repull-after, 13 commit-onto-a-loaded-holder, 14 batch-key-case (B before a), 15 batch-key-punctuation (Z before _ before a), 16 batch-key-case-in-a-group, 17 batch-surrogate-groups (U+1F600 before U+E000, the reverse of code-point order), 18 batch-alias-case (alias Sku before alt), 19 load-alias-case, 20 load-key-case, 21 load-key-punctuation, 22 load-surrogate-groups, 23 batch-holders-by-case (a and A holding one value: A before a), 24 load-holders-by-case. Rows 19-22 and 24 open a store holding the records of rows 18, 14, 15, 17 and 23, so a runtime with no batch commit still walks alias and cross-group key order.

A runtime performs every row its API can perform and asserts the count it walked:

• C# and Rust: all 24.
• Python: 23, every row but 3 (put_envelopes takes one type).
• TypeScript: 12, rows 5, 8, 9, 10, 11, 12, 13, 19, 20, 21, 22 and 24 (no batch commit, so no row 1-4, 6, 7, 14-18 or 23; variants:idea:unique-index-runtimes.ts-batch-commit).