fix(config): complete project language overrides

[andrebrait]

Summary

Complete the project-language override integration from Graphify-Labs#3075 on top of upstream 0.9.59. This PR targets the maintained fork branch, preserving the upstream-tracking v8 branch and all existing tags.

• Apply declared languages inside extractors and cross-file passes, including TypeScript grammar/cache behavior and Python/Java/Go/C# resolution.
• Resolve declared JS/Python module suffixes and package initializers without changing native precedence, physical source paths, or Python 3 import scope.
• Invalidate incremental manifests when language declarations change and handle configuration/source events correctly in the watcher, including relative roots, atomic replacement, and semantic-work notifications.
• Preserve the existing fork's suffix validation and interpreter-gated Leiden packaging.

Verification

Current head: a59563e5c0f63d6a71eaf7df6c4c97ae30564098. The initial changes were independently reviewed at 481574300794093d44ea40d5371db0ef058b21fa; the final commit applies the received Copilot corrections with frozen regressions.

• Full fork suite: 5,720 passed, 14 skipped (existing platform/service-dependent skips).
• Focused regression suite: 124 passed. The initial 99 tests were independently replayed: 39 failed, 60 passed on 9be01cc. Copilot regressions ran 10 failed / 13 passed, plus 2 failed watcher cases before their fixes; the same frozen tests now pass.
• Independent integration gate: 22 additional probes passed, 11 fault mutations detected; no remaining blockers.
• Separate condensed contract/correctness/test-honesty/simplicity review: PASS, with independent red/green and hostile-input replays.
• uv lock --check, Ruff, wheel and source-distribution builds passed. Built production bytes match the reviewed source.
• Real CLI incremental extraction and live watcher processes verified, including default/relative roots and clean shutdown.
• Installed fork provenance verified on Python 3.13.5; native Leiden 1.3.1 successfully partitioned the runtime probe.

Chronological regression evidence and the integration history are recorded in pfBlockerNG/pfBlockerNG#3275. No test assertions were weakened; fixture-only maintenance is explicitly separated from production red/green proof.

Refs pfBlockerNG/pfBlockerNG#3275.

External review disposition

All seven Copilot findings were validated and addressed in the final commit; see the resolution audit and inline reply. CodeRabbit was requested twice and remained unavailable (not reported as a clean review).

[andrebrait]

Independent review and integration gate

Reviewed head: 481574300794093d44ea40d5371db0ef058b21fa (base 9be01cc).

The condensed contract, correctness/hostile-input, test-honesty, and simplicity review passed with no findings remaining. Review invocation: OMP reviewer, anthropic/claude-sonnet-5, requested effort hi. A separate integration verifier independently replayed the regression and mutation evidence.

• Full local suite: 5,695 passed, 14 skipped; skips are documented platform/service cases, not claimed coverage.
• Final focused suite: 99 passed; immutable-baseline replay: 39 failed, 60 passed.
• Independent probes: 22 passed; 11 fault mutations detected.
• Real graphify watch processes for . and src rebuilt graphs, preserved relative source identity, signalled semantic work, and exited cleanly.
• Lock, Ruff, wheel and source-distribution checks passed; built production bytes match the reviewed files.
• Test assertions were not weakened. The recursive no_cluster test-wrapper repair and event-driver extensions are fixture maintenance, not mislabeled production red/green proof. Earlier fix-first drafts were discarded; the accepted integration began with a clean baseline and recorded RED tests.

Per-file disposition

File	Verdict
README.md	PASS — new prose accurately reflects verified behavior: manifest invalidation on declaration change, watch root-config/atomic-replacement handling, JS/Python suffix fallback with native precedence, Fortran/C-header exclusions.
graphify/detect.py	PASS — save_manifest re-activates the given root's overrides before hashing (fixes cross-thread staleness); language-change detection (lang_changed) correctly invalidates ast/semantic tiers independently, handles legacy scalar/dict/tiers rows via prev.get('language') defaulting to None, and never tags never-declared files with a 'language' key (legacy-compatible).
graphify/extract.py	PASS — every dispatch/eligibility/cache-bypass site (_extract_single_file, _extract_sequential, extract, Java/Go/C#/Python cross-file selectors) consistently swapped .suffix for effective_suffix; Python package/sibling repoint passes use the new shared _is_python_package_dir helper with correct default/explicit suffix sets, preserving PEP 328 implicit-relative guard.
graphify/extractors/csharp.py	PASS — _is_cs_file/_is_dotnet_source_file now suffix-aware; extraction to a shared predicate is a genuine DRY improvement with no behavior change for native .cs (case-sensitivity preserved).
graphify/extractors/ocaml.py	PASS — single-line effective_suffix swap for .mli grammar selection; verified via existing mutation-killed test.
graphify/extractors/resolution.py	PASS — largest and highest-risk file; _declared_language_suffixes traversal guard independently verified (see probes above); native-before-declared ordering preserved for both JS file/index resolution and Python module/package resolution; G2/G3/G4 fixes all independently re-verified, not just re-read from logs.
graphify/extractors/robot.py	PASS — .resource grammar selection now suffix-aware; consistent with mutation-killed coverage.
graphify/symbol_resolution.py	PASS — Python-caller eligibility gates now suffix-aware in both collection passes.
graphify/watch.py	PASS — thread-local override snapshot correctly re-seeded on both observer and dispatch threads before any effective_suffix-dependent decision; G1 fix (observer registered on absolute watch_root_for_ignore) verified to keep _rebuild_code(watch_path) on the original — possibly relative — argument, preserving relative source_file identity in the graph, matching gate's real watcher proof (page.tpl / src/page.tpl) and the mocked relative-root test.
tests/test_language_overrides.py	PASS — original 409 lines byte-identical to base (blob 09d3d1d... unchanged across initial/final gate hashes); append-only new regression cases.
tests/test_language_override_lifecycle.py	PASS — new file; assertions are genuine (manifest invalidation, thread-mapped observer/dispatch snapshot, ignored/read-only guards, semantic-flag visibility, relative-root source-path preservation); the reported 'duplicate no_cluster' correction is pure test-harness fixture code (forces no_cluster=True on every rebuild call for test speed) with zero bearing on the production contract under test — correctly test maintenance, not mislabeled production work.
tests/test_language_override_imports.py	PASS — new file; each test fails on the immutable base (independently re-confirmed) and passes on HEAD; includes the traversal-safety regression test I independently reproduced against a second attack surface (Python) not covered by the file's own JS-only traversal test.

No blocking or optional findings remain. Exact-head CI and external review are the remaining landing gates.

[andrebrait]

@coderabbitai review

[andrebrait]

@coderabbitai review

[Copilot]

🟡 Changes recommended

Seven unresolved moderate findings affect override-aware passes, resolution, call-scope handling, and ignored .graphifyrc events.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

Completes project-language override support across extraction, resolution, incremental manifests, and watcher lifecycle.

Changes:

• Applies declared suffixes to extractors and JS/Python resolution.
• Invalidates manifests and handles configuration/source watcher events.
• Adds documentation and regression coverage.

File summaries

File	Summary	Review status
tests/test_language_overrides.py	Tests extractor and cross-file integration.	No final comments.
tests/test_language_override_lifecycle.py	Tests manifests and watcher lifecycle.	No final comments.
tests/test_language_override_imports.py	Tests declared module and package imports.	No final comments.
README.md	Documents override and lifecycle behavior.	No final comments.
graphify/watch.py	Handles configuration and source events.	Moderate (1 vote): process root .graphifyrc events before ignore filtering.
graphify/symbol_resolution.py	Includes declared Python files in symbol resolution.	No final comments.
graphify/extractors/robot.py	Supports declared Robot resource suffixes.	No final comments.
graphify/extractors/resolution.py	Resolves declared JS/Python modules and packages.	Four moderate findings (1 vote each): gate native candidates, package initializers, module paths, and call-scope checks with effective suffixes.
graphify/extractors/ocaml.py	Selects OCaml parsing by effective suffix.	No final comments.
graphify/extractors/csharp.py	Applies overrides to C# detection.	Moderate (1 vote): propagate effective suffixes through partial, inherited, and dispatch checks.
graphify/extract.py	Applies overrides during extraction and cross-file passes.	Moderate (2 votes): replace remaining raw .cs checks in partial merging and base-chain filtering.
graphify/detect.py	Tracks overrides in incremental manifests.	No final comments.

Review details

Suppressed comments (6)

graphify/extractors/csharp.py:153

• Making _is_cs_file() override-aware does not reach all downstream C# consumers. extract.py still filters partial-class nodes and inherited-base edges by physical .cs, while csharp_dispatch.py uses the same raw suffix for both its activation guard and node checks. With language.rzr=csharp, partial types stay split, inherited member calls cannot walk their bases, and interface dispatch exits without adding edges. Propagate the effective-suffix predicate through these passes and add regressions for the declared suffix.

def _is_cs_file(value: object) -> bool:
    return isinstance(value, str) and effective_suffix(value) == ".cs"

graphify/extractors/resolution.py:74

• The native candidate loop checks raw suffixes only, so it can return a file that the override has explicitly changed to another language. For example, with language.ts=php and language.tpl=typescript, an extensionless ./dep resolves to dep.ts here even though that file is no longer JS; the native directory-index loop below has the same problem. Filter native candidates/indexes through effective_suffix before returning so native precedence is preserved only among JS files.

    for ext in _JS_RESOLVE_EXTS + declared_exts:
        with_ext = candidate.parent / f"{candidate.name}{ext}"
        if with_ext.is_file():
            return with_ext

graphify/extractors/resolution.py:2088

• This package-scope check ignores overrides on the native Python suffix. With language.py=javascript (or another non-Python target), __init__.py is not a Python initializer, but this still marks its directory as a Python package and suppresses the sys.path-root walk. Check the initializer with effective_suffix before treating it as a Python package.

    if any((directory / f"__init__{suffix}").is_file() for suffix in suffixes):

graphify/extractors/resolution.py:2115

• The native package probe (and the .py module probe immediately below it) returns raw .py paths without checking whether the project remapped .py away from Python. Thus a Python file using a declared suffix can bind to a non-Python dep.py or dep/__init__.py, contrary to the override contract and the intended native-precedence rule. Gate both native returns with effective_suffix(...) == ".py" before falling back to declared Python suffixes.

    if candidate.is_dir():
        init_path = candidate / "__init__.py"
        if init_path.is_file():
            return init_path
    if candidate.is_file():

graphify/extractors/resolution.py:1802

• By including remapped files in the JS symbol-resolution path here, the PR also sends them through the generic cross-file call pipeline, but that pipeline's no-implicit-scope guard in extract.py still checks rc.source_file.endswith(_JS_TS_CALL_SUFFIXES) against the physical suffix. A language.tpl=typescript caller therefore bypasses the guard and can bind an unimported call to a lone same-named export, unlike a native .ts caller. Use effective_suffix for that raw-call gate and add a declared-suffix no-import regression.

def _collect_js_symbol_resolution_facts(paths: list[Path], facts: _SymbolResolutionFacts) -> None:
    js_paths = [
        path for path in paths
        if effective_suffix(path) in _JS_CACHE_BYPASS_SUFFIXES
    ]

graphify/watch.py:2238

• The watcher applies _is_ignored before checking the root .graphifyrc, so an ignore pattern such as .graphifyrc or .* causes config modify/delete events—and atomic moves whose destination is .graphifyrc—to return here. The old override snapshot then remains active and no rebuild is scheduled. Check the root config (including dest_path) before the ignore filter.

            if ignore_patterns and _is_ignored(path, watch_root_for_ignore, ignore_patterns):

• Files reviewed: 12/12 changed files
• Comments generated: 1
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[andrebrait]

Applied in a59563e. Shared C# predicates now cover partial merging, inherited-member base chains, scoped rewiring and interface dispatch. Native/declared regressions failed before the fix and pass unchanged afterward; full suite: 5,720 passed, 14 existing skips.

[andrebrait]

Applied in a59563e5c0f63d6a71eaf7df6c4c97ae30564098.

Received finding	Resolution
C# partial/inherited passes	Reused effective-language C# predicates in partial merging, inherited-base filtering and scoped stub rewiring.
C# interface dispatch	Reused the same predicate for corpus and endpoint eligibility.
Native JS candidates remapped away	Native file, index and ESM-substitution candidates must still be JS-family; declared alternatives retain their fallback role.
Native Python package markers remapped away	Package markers must retain an effective Python suffix.
Native Python module/package candidates remapped away	Native .py candidates are checked before returning; valid declared Python alternatives remain available.
JS no-implicit-call scope	Declared JS/TS callers now pass through the existing import-evidence guard.
Ignored root configuration	Root .graphifyrc control events, including atomic replacement, are handled before corpus ignore filtering. Other ignore/read-only/output guards remain unchanged.

Targeted tests ran before production edits: 10 failed / 13 passed, plus 2 failed watcher cases. The same frozen tests now pass; the combined focused set is 124 passed. Full suite: 5,720 passed / 14 existing skips. Ruff, lock and package builds passed.

These are the review's concrete corrections, not a new investigation or feature pass. Per the concrete-suggestion fix-round exception, no additional broad adversarial round was opened. Exact-head CI is being rerun.

CodeRabbit was requested twice on this fork PR and produced no response. This is recorded as unavailable, not a clean review; the missed-review ledger will reference the landed SHA.

[andrebrait]

Landed by signed fast-forward of the maintained fork branch at a59563e5c0f63d6a71eaf7df6c4c97ae30564098; post-push PR head fence and ancestry checks passed. GitHub reports this PR MERGED. Signed tag v0.9.59-pfb.1 resolves to that exact commit. Exact-head CI is green on Python 3.10/3.12, skill-generation and security jobs. All Copilot findings were addressed; CodeRabbit was silent after both permitted requests. Existing tags and the upstream-tracking branch were not rewritten.