Enable three new mypy strictness flags and run pyright as an additional type checker, fixing all reported errors across the codebase.

What do these changes do

• Adds extra_checks, strict_bytes, strict_equality_for_none, and strict_optional to [tool.mypy] in pyproject.toml.
• Adds pyrightconfig.json (basic mode) that suppresses false positives from uninstalled optional dependencies (torch, numpy, transformers, etc.) while keeping real checks enabled.
• Fixes ~20 type errors surfaced by pyright across 13 source files.

What was wrong

mypy was missing three strictness flags. pyright had no config and had never been run — it surfaced:

• list[LiteralString] assigned to list[str] (invariant list) in tcc.py, tcc_p.py, pronounce.py
• Possibly-unbound variables: attn_energies in thai2rom/thaig2p (no else branch on if/elif chains), th_m/days_in_month in thai_lunar_date
• Global lazy-load pattern (Optional[X] globals) not narrowed after assignment — unigram.py tagger functions
• Duplicate/conflicting type re-annotations in _tag_perceptron.py load() (re-declared classes as list[str] then called .add() on it), tokenization_small100.py, translate/
• Broad NER.tag() return union not narrowed before tuple unpack in nercut.py
• Trie | None passed where Trie | list[str] | str expected in tokenize/core.py

How this fixes it

• pyproject.toml: three new mypy flags added in alphabetical order among existing flags.
• pyrightconfig.json: suppresses reportMissingImports, reportAttributeAccessIssue, reportCallIssue, reportGeneralTypeIssues, reportInvalidTypeForm, reportOperatorIssue, reportOptionalMemberAccess — all noise from optional deps absent in CI; mypy already covers these via ignore_missing_imports overrides.
• _tag_perceptron.py: removed redundant class-body declarations; fixed load() to assign set(w_td_c["classes"]) (not retype as list[str]); replaced prev, prev2 = self.START with indexed access.
• unigram.py: cast(dict[str, str], _ORCHID_TAGGER) at return site — standard pattern for lazy-init globals pyright can't narrow.
• thai2rom.py / thaig2p.py: added else: raise ValueError("Unsupported attention method: …") to make attn_energies always bound.
• thai_lunar_date.py: th_m = 0 before loop; else: raise ValueError for unknown last_day.
• tokenization_small100.py: removed instance-level re-annotations for already-declared attributes; # pyright: ignore[reportRedeclaration] on property getter/setter pair (false positive from unresolvable transformers base class).
• budoux.py: explicit None guard after lazy init so pyright narrows Optional[Any] → non-None before .parse().

Your checklist for this pull request

• Passed code styles and structures
• Passed code linting checks and unit test

📍 Connect Copilot coding agent with Jira, Azure Boards or Linear to delegate work to Copilot in one click without leaving your project management tool.