Harden parser, speed up graph ops, add incremental cache

[LeTuR]

Summary

Correctness, performance, design, and robustness improvements across the analyzer, with 36 new regression tests (176 total, ruff clean).

Parser correctness

• Strip comments and triple-quoted strings before matching preload/load/extends — commented-out or in-string references no longer create phantom graph edges (which could manufacture false cycles/rule violations).
• Expanded the builtin class set to cut false "broken inheritance" edges.

Performance

• get_dependents now uses a lazily-built reverse-adjacency index (was O(V·E) per call, quadratic in the orphan check).
• Scanner walks the project tree once, bucketing by suffix (was 3 full walks).
• Tarjan cycle detection converted to an iterative stack — no RecursionError on deep chains.

Design

• Shared type_label/short_path/to_res_path helpers; deleted the boilerplate Patterns class; formatter registry; DOT/Mermaid label escaping.

Features & robustness

• Incremental parse cache keyed by mtime+size (--cache / --cache-file); symbol resolution and cycle detection always run fresh so output is identical to an uncached run.
• Warnings for unknown config keys and bad severity values.
• Duplicate class_name collisions surfaced as warnings; malformed project.godot and non-UTF-8 files no longer abort the run.

Test plan

• uv run pytest → 176 passed (36 new: parser edge cases, cache hit/miss/invalidation, dependents index, deep-graph cycles, scanner, output escaping, symbol collisions, malformed input, config semantics).
• uv run ruff check . → clean.
• Verified manually: phantom deps gone / real deps kept, deep-chain Tarjan, cache warm-run equals cold-run graph, single-file cache invalidation, config warnings surface on typo'd sections.