FIX: Preserve word-selection parameters in converter identifiers

[Zhibo Lin (LE0-Lin)]

Description

Fixes #2995. Different word selections can produce different text but the same converter identifier, causing the second default registry registration to fail.

I extended the BinAscii fix to CharSwap, FirstLetter, Leetspeak, UnicodeReplacement, and Zalgo. They now include the parent word-selection parameters. Existing default identities are preserved, including CharSwap's default 20% selection and the other converters' all-words configurations. Non-default selections intentionally receive corrected hashes; previously saved hashes for those configurations can differ.

WordIndexSelectionStrategy also retains the supplied index order in its identity. Reordering indices can change which random values are used for each word. The fix preserves existing conversion order and outputs rather than sorting execution.

This follows the maintainer invitation to continue in this PR. Constructor APIs and conversion methods are unchanged.

Tests and Documentation

• Original BinAscii regressions: 5 failed / 7 passed before its fix. The other five converter overrides: 25 failed / 15 passed before the expansion.
• Updated index-order regressions: 9 failed against the previous production code. New seeded CharSwap, Zalgo, and nondeterministic Leetspeak cases check different but repeatable outputs, distinct identities, and successful registration of both instances. Independent before/after runs confirm conversion outputs are unchanged.
• Related converter, word-selection, seeded-determinism, and registry tests across 12 modules: 389 passed. Python 3.12.0, frozen lockfile, full-extras environment.
• uv run --frozen --all-extras ty check on the five files changed for review: passed, including the test files. The fixture now narrows request.param before returning it.
• Normal repository pre-commit run --files on those five files: passed, including full-tree ty check pyrit. git diff --check passed.

The WordIndex selection docstrings now describe conversion order. Full repository tests and real-model integrations were not run locally.

AI assistance was used.

[Zhibo Lin (LE0-Lin)]

I included the other five converter overrides and kept the legacy default hashes. The related 321 tests and pre-commit checks pass locally. This is now ready for review, Roman Lutz (@romanlutz). Thanks!

[Roman Lutz (romanlutz)]

🔴 Must Fix: reordered indices still collide for seeded converters. These index lists are not equivalent for all of the converters in this test. With CharSwapConverter(max_iterations=1, seed=123) and prompt "alpha beta", [0, 1] produces "alpah btea", while [1, 0] produces "alhpa btea". Both results are repeatable, but the hashes and default registry names are identical, so registering the second still raises ValueError.

WordIndexSelectionStrategy.get_identifier_params() sorts the indices, while select_words() returns them in their original order. CharSwap consumes one shared random stream in that order. I also reproduced this with seeded Zalgo and nondeterministic Leetspeak under a fixed root seed.

Could we add actual conversion assertions for these stochastic cases and preserve behavior-changing order in the identity, or explicitly make execution order canonical? Hash equality alone hides the remaining collision.

[Zhibo Lin (LE0-Lin)]

You're right, I missed the effect of execution order on the random stream. I now keep that order in the identity and added conversion and registration tests for all three cases. The outputs are unchanged by the fix.

[Roman Lutz (romanlutz)]

🟡 Should Fix: this fixture fails the configured test type check. uv run --frozen ty check tests\unit\converter\test_word_selection_identifiers.py reports unsound-return-statement here because request.param is Any, not a verified type[WordLevelConverter]. The pre-commit hook checks only pyrit, so it does not catch this test-file error.

Could we narrow the value before returning it? This version passes the same type-check rules:

converter = request.param
assert isinstance(converter, type)
assert issubclass(converter, WordLevelConverter)
return converter

[Zhibo Lin (LE0-Lin)]

Added the narrowing assertions you suggested. ty now passes for all five changed files, including the tests.