docs: add core runtime benchmark harness notes

Author: jdrhyne

.gitignore

@@ -21,6 +21,7 @@ node_modules/
 
 # Build output
 dist
+.autoresearch-dist/
 
 # Temporary files
 tests/assets/test-output-*

autoresearch.ideas.md

@@ -0,0 +1,5 @@
+- Revisit `src/fs/directoryTree.ts` with a different strategy than fully sequential traversal: e.g. limited-concurrency batching for readability checks, or a cheaper readability probe than `open()` if test expectations can be updated safely.
+- Explore a more selective directory-tree cache invalidation strategy keyed by subtree or filesystem metadata so repeated tree reads stay hot without over-invalidating on unrelated writes.
+- Try sandbox-mode function-pointer specialization for `resolveReadFilePath()` / `resolveWriteFilePath()`: switch a small exported wrapper between sandbox-disabled and sandbox-enabled implementations inside `setSandboxDirectory()` so the hot path can avoid repeating the nullable active-cache branch on every cached hit.
+- Revisit write-path epoch invalidation with a single cache-status table instead of extra `Set` lookups: encode whether a validated writable path was an existing file or only a creatable target, then skip directory-tree epoch bumps only for the existing-file case to reduce the huge `writes-existing` hotspot without paying the branchy cost of multiple membership checks.
+- Add a narrow microbenchmark for cached read/write wrapper hits only, so wrapper-shape candidates can be screened quickly before paying for the full x800/51 harness; the full benchmark can remain the final authority, but a cheaper prefilter may let more promising sandbox-wrapper ideas be explored.