Slice: Pinned Unicode 16.0.0 tables, the Python library, vector generation Plan: identity-minting.md Status: identity-minting.md Mode: autonomous (granted 2026-09-23)

What invariants does this slice protect?

The library mints what the anchors say. Every anchor vector, verified independently with openssl, passes trace step by trace step.

The committed vectors are what the library generates today. A behaviour change cannot land without regenerating packages/minting/testdata and showing the diff.

Unicode behaviour comes from the pinned tables, and the tables are right. Both libraries carry byte-identical tables, the tables match the UCD 16.0.0 files, and over every code point they agree with CPython’s own 16.0.0 data wherever the two define the same thing.

Every refusal is named. Recipe faults, runtime faults and request faults each raise their §9 error.

The library stands alone. No dependency outside the Python standard library, and no import of any other LATTICE package.

Test case table

ID	Given / When / Then	Level	Pass criterion	+/-
T1	All seven anchor sets	L3	`test_conformance.py::test_anchors_pass`	+/-
T2	All seven generated vectors files	L3	`::test_generated_vectors_pass` (7)	+/-
T3	Regenerating each file’s vectors from its embedded recipe gives the committed document	L2	`::test_generated_vectors_are_current` (7)	+
T4	Each `.recipe.json` equals the recipe embedded in its vectors, and parses	L1	`::test_recipe_files_match_vectors` (7)	+
T5	The compiled recipes are the anchor recipes	L2	`::test_anchor_recipes_equal_compiled_recipes`	+
T6	A key of default-ignorables only is refused under `NfkcTrimCasefold` and mints under `NfkcTrimUppercase`	L1	`::test_default_ignorable_only_depends_on_pipeline`	+/-
T7	A changed expected IRI, and a changed trace step, are reported, the second by step number	L1	`::test_a_changed_expectation_is_reported`, `::test_a_changed_trace_step_names_the_step`	-
T8	`lattice-mint verify` exits 0, 2 on bad usage, 1 on a failing file	L4	`::test_cli`	+/-
T9	Python and Java table copies are byte-identical	L1	`test_unicode.py::test_both_libraries_carry_identical_tables`	+
T10	Assigned set equals General_Category not `Cn`, over all 1,114,112 code points	L2	`::test_assigned_matches_general_category`	+
T11	Full upper and lower mappings equal `str.upper` and `str.lower` for every assigned code point, and `Final_Sigma` in context	L2	`::test_full_case_mappings_match_python`, `::test_final_sigma_matches_python` (7)	+
T12	`nfkc_casefold` equals the NFKC and case folding composition where it keeps a character, and removes only format, mark, letter-other or control characters	L2	`::test_nfkc_casefold_agrees_with_the_casefold_composition`	+
T13	`White_Space` differs from `str.isspace` only at U+001C–U+001F	L2	`::test_white_space_differs_from_isspace_only_at_the_information_separators`	+
T14	All three pipelines are idempotent on every assigned code point	L2	`::test_every_pipeline_is_idempotent`	+
T15	Wrong format, unknown strategy, changed content, another Unicode version, claims with different keys	L1	`test_refusals.py` (5 cases)	-
T16	Digest ignores whitespace and member order	L1	`::test_whitespace_and_member_order_do_not_affect_the_digest`	+
T17	Unicode data older than 16.0.0 is refused	L1	`::test_older_runtime_unicode_is_refused`	-
T18	Error kinds are closed, random surrogates use the supplied source and are well formed, secrets are copied	L1	`::test_error_kinds_are_closed`, `::test_random_surrogates_use_the_supplied_source`, `::test_minter_does_not_share_secrets_with_the_caller`	+/-

Given / When / Then

Level

Pass criterion

+/-

All seven anchor sets

test_conformance.py::test_anchors_pass

+/-

All seven generated vectors files

::test_generated_vectors_pass (7)

+/-

Regenerating each file’s vectors from its embedded recipe gives the committed document

::test_generated_vectors_are_current (7)

Each .recipe.json equals the recipe embedded in its vectors, and parses

::test_recipe_files_match_vectors (7)

The compiled recipes are the anchor recipes

::test_anchor_recipes_equal_compiled_recipes

A key of default-ignorables only is refused under NfkcTrimCasefold and mints under NfkcTrimUppercase

::test_default_ignorable_only_depends_on_pipeline

+/-

A changed expected IRI, and a changed trace step, are reported, the second by step number

::test_a_changed_expectation_is_reported, ::test_a_changed_trace_step_names_the_step

lattice-mint verify exits 0, 2 on bad usage, 1 on a failing file

::test_cli

+/-

Python and Java table copies are byte-identical

test_unicode.py::test_both_libraries_carry_identical_tables

T10

Assigned set equals General_Category not Cn, over all 1,114,112 code points

::test_assigned_matches_general_category

T11

Full upper and lower mappings equal str.upper and str.lower for every assigned code point, and Final_Sigma in context

::test_full_case_mappings_match_python, ::test_final_sigma_matches_python (7)

T12

nfkc_casefold equals the NFKC and case folding composition where it keeps a character, and removes only format, mark, letter-other or control characters

::test_nfkc_casefold_agrees_with_the_casefold_composition

T13

White_Space differs from str.isspace only at U+001C–U+001F

::test_white_space_differs_from_isspace_only_at_the_information_separators

T14

All three pipelines are idempotent on every assigned code point

::test_every_pipeline_is_idempotent

T15

Wrong format, unknown strategy, changed content, another Unicode version, claims with different keys

test_refusals.py (5 cases)

T16

Digest ignores whitespace and member order

::test_whitespace_and_member_order_do_not_affect_the_digest

T17

Unicode data older than 16.0.0 is refused

::test_older_runtime_unicode_is_refused

T18

Error kinds are closed, random surrogates use the supplied source and are well formed, secrets are copied

::test_error_kinds_are_closed, ::test_random_surrogates_use_the_supplied_source, ::test_minter_does_not_share_secrets_with_the_caller

+/-

Commands

mise run bootstrap:minting-python
mise run check:minting-python     # 51 passed
mise run check:minting-anchors    # 187 checks passed
mise run check:minting-tables     # 16 tables current
mise run check:persistence        # 629 passed

All four green on 2026-09-24, Python 3.14.7 (Unicode data 16.0.0).

Adversarial probes

Probe	Result
`Final_Sigma` disabled	5 failures
`str.strip()` in place of `trim_white_space`	8 failures
Truncation keeps the last bytes	8 failures
`NFKC(casefold(s))` in place of `nfkc_casefold`	5 failures

Probe

Result

Final_Sigma disabled

5 failures

str.strip() in place of trim_white_space

8 failures

Truncation keeps the last bytes

8 failures

NFKC(casefold(s)) in place of nfkc_casefold

5 failures

All restored, suite green.

Contract changes (non-weakening rule)

Generated vector files escape non-ASCII. export-recipes --vectors writes JSON with \u escapes, and anchor-vectors.json was rewritten the same way, so invisible characters in inputs can be read in review. The parsed content is unchanged (the anchor verifier and both suites pass unchanged).

A recipe’s claims must share one key (UnsupportedRecipeFormat otherwise, specification §6.3). The compiler has always emitted them that way.

Specification brought up to the library: request inputs table (§6), randomHex, validate-token, generate-surrogate, validate-iri, CanonicalizerNotDeclared, which faults raise MissingKeyComponent, truncate present even at full width, the omitSecret harness directive, and the test secret derivation (§10.1).

Found on the way

Default-ignorable code points under the upper- and lowercase pipelines. NFKC keeps zero-width space and soft hyphen, and only nfkc_casefold removes them. A key made of nothing else is refused as empty under NfkcTrimCasefold but mints under the other two pipelines. The generator now emits the vector each pipeline actually implies. Whether those pipelines should remove default-ignorables is a design question, recorded in the status record.

Tool input turned \u escapes into real characters. Invisible characters reached vectors.py. Every non-ASCII literal in the library source is now written with chr(), and all changed files were scanned for format, control and non-ASCII space characters.