fix(cache): scope semantic cache writes to extracted files

[TPAteeq]

Summary

• Add an allowed_source_files boundary to save_semantic_cache.
• Prevent semantic results from replacing cache entries for files outside the dispatched extraction batch.
• Apply the boundary to both native CLI extraction and every generated agent workflow.
• Preserve out-of-scope nodes in the graph while skipping and warning about unsafe cache writes.
• Preserve backward compatibility for callers without an allowlist.

Validation

• 3117 passed, 30 skipped
• ruff check passed
• tools.skillgen --check passed
• Monolith round-trip validation passed
• graphify update . completed successfully

Closes #1757

[safishamsi]

Landed on v8 (579ba1d, your commit preserved) plus a follow-up (da9616d) — ships in 0.9.14. Thanks @TPAteeq, the allowlist design is right and the skillgen threading was clean.

One gap I fixed on top: the guard covered the final CLI write, but the per-chunk incremental checkpoint in llm.py:_checkpoint_chunk — the write that actually runs on every extract/update via extract_corpus_parallel — was still calling save_semantic_cache without an allowlist (merge_existing=True), so a stray mis-attributed node could still pollute a victim file's entry before the final write. I scoped that call to the chunk's own dispatched files and added an extract_corpus_parallel integration test proving an out-of-chunk attribution leaves the victim's cache untouched. Also hoisted the import warnings to module level. Credited your commit in the history.