Add multi-codepoint confusable support to homoglyph search

[codeboxbot]

Summary

Adds support for multi-codepoint confusables — cases where a single glyph in the text stands in for a sequence of characters in the target word. For example the ligature ffi (U+FB03) resembles ffi, so office is now found when searching for office; the letter m resembles rn, so comer is found when searching for corner.

Previously the generator's data-file parser called chr(int(code, 16)) on the target field, which raised on any multi-codepoint target (e.g. "0066 0066 0069") and silently dropped it, so all ~2000 multi-codepoint confusables in the UTS #39 data were discarded. This PR captures them via a separate mapping and threads them through every consumer.

This is separate from PR #21 (the UTS #39 v17 data refresh) and does not touch it.

Semantics

When matching a target word against the text, a single text symbol matches either:

• one word character (existing behaviour: exact match or single-codepoint homoglyph), or
• a run of word characters, when that symbol is a multi-codepoint confusable whose prototype sequence equals that run (case-insensitively).

Alternatives are explored with backtracking, so a multi-codepoint substitution composes with ordinary single-codepoint substitution (e.g. 0ffice → office, where 0→o and ffi→ffi). Single-codepoint behaviour and the public search(...) APIs are unchanged; the matched substring may now be shorter than the target word.

Changes

Generator

• data_file_parser.py: detects multi-codepoint targets (is_multi/get_multi) and routes them to a new manager instead of dropping them.
• multi_char_manager.py (new): collects glyph → prototype-sequence mappings.
• output_multi_char_codes.py (new) + templates/multi_char_codes.txt (new): emit raw_data/multi_char_codes.txt.
• output_js.py: emits a second MULTI_CHAR_MAP into the JS artifacts and now sorts the single-codepoint glyph lists so regeneration is deterministic.
• main.py: wires the new manager and output in.

Runtime

• raw_data/multi_char_codes.txt (new): <glyph hex>;<space-separated prototype hex> per line.
• javascript/src/homoglyph.js, node/index.js: regenerated — new backtracking matcher + embedded MULTI_CHAR_MAP. buildSearchFunction takes an optional second multiCharMap argument (defaults to {}, so existing callers are unaffected).
• Java Homoglyph.java: new backtracking matcher; new Homoglyph(List<Set<Integer>>, Map<Integer,List<int[]>>) constructor (old constructor preserved). HomoglyphBuilder.java: loads the optional bundled multi_char_codes.txt; new build(Reader, Reader) overload. build.gradle: bundles the new resource.

Tests

• generator/tests/test_multi_codepoint.py (new): parser/manager/output regressions using real UTS #39 lines — 13 tests.
• node/test/spec/MultiCharTests.js + javascript/tests/js/tests/MultiCharTests.js (new): end-to-end JS regressions — 15 specs.
• src/test/java/net/codebox/homoglyph/HomoglyphMultiCodepointTest.java (new): end-to-end Java regressions — 12 tests.

Test results (all run locally)

• Python generator suite: 13/13 pass.
• Java (HomoglyphLogicTest + HomoglyphDataTest + HomoglyphMultiCodepointTest, Temurin JDK 17, JUnit 4.11): 24/24 pass (12 pre-existing + 12 new).
• Node (npm test): 15 new multi-codepoint specs pass. Total 1579 specs, 32 failures — the 32 failures are pre-existing on upstream master (the hand-maintained node/test/spec/DataTests.js has drifted from the bundled data; unrelated to this change).
• raw_data/char_codes.txt and raw_data/chars.txt are byte-identical to master (single-codepoint data unchanged).

Caveats

• Scope of the raw multi data: only confusables whose prototype is composed of ordinary word characters ([A-Za-z0-9.-]) are emitted (124 mappings), matching how the JS single-codepoint map is already filtered. Combining-mark and other non-word prototype sequences are excluded because they can never occur in a searchable target word. (Java therefore supports multi-codepoint detection for word-character prototypes only; single-codepoint detection is unchanged and still uses the full data.)
• Determinism: the single-codepoint JS glyph lists are now emitted in sorted order so python generator/main.py is reproducible; this accounts for the reordering in the homoglyph.js/node/index.js diffs (membership is unchanged).
• Performance: matching is now recursive with backtracking. Word lengths are tiny and the multi-map is only consulted when a text symbol is itself a multi-codepoint confusable, so the impact is negligible; worst case is bounded by target-word length.
• The pre-existing node/test/spec/DataTests.js drift is left untouched (out of scope).