Perl: expose file_filter/ignore_dirs via ls_specific_settings (#1449)

[weiconghe]

Fixes #1449

What & why

Perl::LanguageServer only indexes files whose extension is in its perl.fileFilter and skips directories in its perl.ignoreDirs. Both lists were hardcoded in PerlLanguageServer, so projects whose Perl source uses non-standard extensions — e.g. .cgi / .psgi web handlers, which can outnumber .pm/.pl in mature Perl backends (the reporter's example: 825 .cgi vs 340 .pm + 285 .pl) — had no way to make those files visible to find_symbol / find_referencing_symbols without forking Serena.

860c5841 added .t to the default fileFilter as a surface-level fix (one more extension). This PR addresses the root configurability gap that the maintainer acknowledged in #1449:

MischaPanch: "The proposed patch works, would you like to open a PR with it?"

The reporter's proposed approach is exactly what's implemented here. Opening it myself since the issue has been idle for ~2 months after that invitation, with no PR or follow-up from the reporter. All diagnosis credit for the configurability direction goes to @aaronbamblett.

Change

Surface the two lists through ls_specific_settings["perl"], mirroring the established pattern in groovy_language_server.py:

ls_specific_settings:
  perl:
    file_filter: [".pm", ".pl", ".t", ".cgi", ".psgi"]
    ignore_dirs: [".git", "blib", "local", "vendor", "cover_db"]

• PerlLanguageServer.__init__ now resolves self._file_filter / self._ignore_dirs from solidlsp_settings.get_ls_specific_settings(Language.PERL), defaulting to the existing values, so behaviour is unchanged for projects that don't set the override.
• Both LSP handshake sites (workspace/configuration handler and didChangeConfiguration notification) read the instance attributes instead of the old class constant / inline literals.
• The old _FILE_FILTER class constant is replaced by module-level _DEFAULT_FILE_FILTER / _DEFAULT_IGNORE_DIRS, with the "keep in sync with Language.PERL.get_source_fn_matcher()" note preserved on the default.

This follows serena's CONTRIBUTING.md scope ("small bug fixes" / "extend along existing lines" via PR without prior discussion), and the trusted-project gating added recently to ls_specific_settings applies automatically.

Verification

Added test/solidlsp/perl/test_perl_config.py — 6 unit tests against the pure _resolve_filter_settings helper, deliberately not requiring a Perl runtime (existing Perl integration tests are skipped when Perl::LanguageServer isn't installed; these run everywhere):

• defaults returned when ls_specific_settings is absent
• defaults returned when only another language is configured (no cross-language leak)
• custom file_filter makes extra extensions (.cgi, .psgi) visible — the Perl: file extensions hardcoded to .pm/.pl — exclude legitimate Perl source like .cgi #1449 scenario
• custom ignore_dirs applied
• both overrides apply independently
• default module-level lists aren't mutated across instances (the resolver copies defaults)

Local checks on Windows / Python 3.13:

$ uv run --extra dev pytest test/solidlsp/perl/test_perl_config.py -v
6 passed
$ ruff check src/solidlsp/language_servers/perl_language_server.py test/solidlsp/perl/test_perl_config.py
All checks passed!
$ ruff format --check <same files>
2 files already formatted
$ ty check src/solidlsp        # _ty_core
All checks passed!
$ ty check test --exclude test/resources   # _ty_test
All checks passed!

CHANGELOG.md updated under "Language Servers".

[opcode81]

The problem is that Serena's symbol index is exclusively determined by the FilenameMatcher returned by Language.PERL.get_source_fn_matcher(). If the matcher is out of sync with what is passed to the language server, this leaves you with inconsistent behaviour:

• The language server will return symbols for a file with a different extension (e.g. a .cgi file) when explicitly asked to do so.
• But find_symbol will not find symbols that don't use one of the extensions in the FilenameMatcher.

Therefore, the out-of-sync case is a problem, and changing the extensions in only one place won't give you the desired behaviour. We would need to modify the FilenameMatcher as well, and we could easily do so if we turn the instances into per-language singletons.

[weiconghe]

Good catch — you're right that find_symbol relies on Language.PERL.get_source_fn_matcher(), so extending only the LS-side fileFilter would leave the matcher (and thus the symbol index) out of sync. Fixed in the updated push.

What I did:

• Added FilenameMatcher.add_extensions(*exts) (idempotent, respects case_sensitive) — a small mutator intended for the now-@cached per-language singletons.
• In PerlLanguageServer, after resolving file_filter from ls_specific_settings["perl"], I call Language.PERL.get_source_fn_matcher().add_extensions(*file_filter), so the configured extensions reach every consumer of the matcher: project.is_ignored_path (symbol index traversal), ls.is_ignored_path, and the language-composition detection.

The default file_filter ([".pm", ".pl", ".t"]) matches the existing Perl matcher, so behaviour is unchanged for projects that don't set the override; only added extensions (e.g. .cgi) propagate. The sync is a separate _sync_source_fn_matcher step so _resolve_filter_settings stays a pure, unit-testable function.

Tests added (test/solidlsp/perl/test_perl_config.py): the resolver cases, FilenameMatcher.add_extensions (new extension, idempotency across repeated calls, case-insensitive normalisation), and the matcher sync (custom file_filter extends the Perl matcher; default file_filter leaves it unchanged — both restore the singleton afterwards). 11 tests pass; ruff/ty clean.

[opcode81]

There are further subtleties to consider: If a new project is activated following the activation of a project that reconfigured the file extensions, then the original extensions must apply; the reconfiguration must be undone.
Therefore,

• the matcher needs a reset method (and it needs to store the original configuration for this work)
• the reset method must be called upon language server initialization. The natural place for this would be in SolidLanguageServer.__init__ after self.language = self.get_language_enum_instance().

[weiconghe]

Good catch on the cross-project leak — that's a real defect in my add-only approach. Once project A had added .cgi, the singleton would have kept it for every subsequently activated project. Thanks for spotting it.

Implemented exactly as you suggested:

• FilenameMatcher now snapshots its initial extensions at construction and exposes reset(), which restores that snapshot.
• SolidLanguageServer.__init__ calls self.language.get_source_fn_matcher().reset() right after self.language is determined, so each activation starts from the language's defaults before the subclass re-applies its own file_filter (Perl's _sync_source_fn_matcher runs after super().__init__(), so the ordering is reset → re-apply).
• Added test_reset_prevents_cross_project_leak in test_perl_config.py, which reproduces the scenario directly: project A syncs .cgi, project B activates (reset + default config), and asserts .cgi is gone while .pm survives.

15 tests pass; ruff / ty clean. Pushed as de4b78ed.

[weiconghe]

Thanks — moved the reset below the self.language attribute docstring so it stays attached to the attribute it documents. Pushed as 30bb6b87.

[weiconghe]

There was no Perl section yet, so I added #### Perl in lexicographical position (between Pascal and PHP). It documents both file_filter and ignore_dirs under ls_specific_settings.perl — defaults, an example serena_config.yml snippet, and notes that configured extensions are also synced into Serena's source-file matcher (and reset on each activation so they don't leak across projects). Pushed as 967472e4.