perf: avoid unnecessary work in file dependency checks

[TristanKruse]

Summary

Speed up direct file-dependency checks without changing their results:

• Cache source/target filter results per distinct path label for the duration of one rule check.
• Inventory all project files for correct internal-import resolution, but parse imports only from files selected by a direct rule's source filters. Cache per-file edges within the process so later rules and full-graph checks reuse them. Empty source filters still request the full graph, and cache clearing invalidates both caches.
• For glob filters used with search(), avoid matching a redundant leading * prefix. The original compiled regex remains available on Filter.regexp; only the internal search path uses the equivalent faster form. Custom compiled regexes are unchanged.
• Walk each parsed syntax tree once, then reuse the node sequence to classify conditional/type-checking imports and collect imports.
• Skip per-file evaluation of default exclusions that cannot match a .py filename; directory pruning, custom file exclusions, and .archignore still apply.

Refs #58 and #59. The reusable benchmark work remains tracked in #55.

Verification

On a generated project with four areas and a direct api → infrastructure prohibition, I alternated clean 1.7.0 and commit 2f7efbf in eight fresh-process pairs per size. Both versions passed the clean, seeded forbidden-import, and cycle correctness gates before timing:

Leaf modules	1.7.0 median	2f7efbf median	Change	Faster pairs
100	0.831 s	0.784 s	-5.6%	6/8
1,000	1.357 s	0.878 s	-35.3%	8/8
5,000	3.559 s	1.349 s	-62.1%	8/8

The final exclusion change was then measured independently against 2f7efbf, again in eight fresh-process pairs with the same correctness gates:

Leaf modules	2f7efbf median	Current branch median	Change
100	0.790 s	0.778 s	-1.5%
1,000	0.903 s	0.838 s	-7.2%
5,000	1.391 s	1.206 s	-13.3%

These are local Windows 11 / Python 3.13.4 results for one synthetic direct-boundary rule, not a universal ranking. At 5,000 leaf modules, selective extraction parses 1,251 of the 5,005 project files. The two tables are separate paired experiments; their absolute times should not be combined across runs.

• Local full suite: 313 passed on Python 3.12 and 3.13; Ruff and Mypy pass.
• Tests cover partial/full graph equivalence in both call orders, namespace-package aliases, import-kind merging, type-only/conditional/dynamic imports, combined source filters, empty filters, no matches, explicit exclusions, .archignore, cache invalidation, and preservation of glob search and Filter identity semantics.
• Full graph edge counts and hashes matched clean 1.7.0 on the generated 1,000-module project and the sample-project fixture.

Change Type

• Refactoring