docs(changelog): note tier-1 bugfixes, perf, and quality work

Author: raulk

CHANGELOG.md

@@ -6,6 +6,21 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/).
 
 ## [Unreleased]
 
+### Added
+
+- `omnifs dev` and `omnifs up` now thread the session's preopened provider paths into the runtime container launch, so providers that require host-path access (such as the db provider's SQLite database directory) are bound into the container.
+
+### Changed
+
+- Lower latency and memory churn on the read and listing hot paths: `read_file` now does a single canonical-object lookup per call, a projection that materializes many objects commits its object-cache writes in one redb transaction instead of one fsync per object, and object leaf-set merges and non-exhaustive directory-listing merges are no longer quadratic in the size of the object or directory. Observable results are unchanged.
+- Internal code-quality pass across the host, SDK, CLI, and providers. No behavior change.
+
+### Fixed
+
+- FUSE now preserves a learned exact file size across later cached-metadata refreshes. A full-deferred read can learn a file's size before a refresh revisits the same inode; the refresh previously reset the size to the unknown-size sentinel, so stat-based tools (`ls -l`, `wc -c`, `du`) reported an unknown size after the content had already been materialized. The learned size is now kept when the refresh still describes the same content identity.
+- Negative-lookup ("not found") cache entries no longer grow without bound. Tombstones were garbage-collected, but expired negative entries lingered until their exact path was read again or invalidated, accumulating indefinitely on long-running mounts with many failed lookups. They are now swept alongside the tombstone GC.
+- The arXiv provider no longer aborts on a JSON encode failure. Pretty-JSON serialization is now a single shared SDK helper with one non-panicking failure policy.
+
 ## [0.2.1] - 2026-06-08
 
 ### Fixed