Summary

Cuts the v0.1.35 release. Three-file diff (per the split-PR pattern):

• packages/memtomem/pyproject.toml — version = "0.1.35"
• uv.lock — workspace pin sync
• CHANGELOG.md — [Unreleased] → [0.1.35] — 2026-05-02, with the
  merged-since-v0.1.34 entries folded in by section

What's in this release

Added

• Model-readiness banner for mm web (feat(web): surface fastembed model load state via banner (#696) #703, closes First search is silent for 30s+ while embedding/reranker models download — surface progress #696). New
  GET /api/system/model-readiness endpoint reports per-component
  load state (embedder + reranker) derived from _loading /
  _load_error flags plus a fastembed cache filesystem probe. The
  header banner polls the endpoint and renders
  Downloading bge-m3 (~2300 MB)… / Loading model… /
  Model failed to load so the first search after a cold-cache boot
  no longer feels like a hung UI.
• i18n wired up across Export / Import / Tags sort / Auto-Tag /
  Timeline (feat(web): wire up missing data-i18n on Export, Tags sort, Auto-Tag, Timeline #695) and the Sources "Load All" button (i18n: promote "Load All" button label to chunks.load_all locale key #727, closes
  i18n: promote "Load All" button label to chunks.load_all locale key #681) — Korean-mode users no longer see Source Filter, Today,
  Last 7 days, Count↓, Dry run, Limit, etc. as untranslated
  English mid-page.

Fixed

• mm web SPA + /api/docs work fully offline (fix(web): vendor cdnjs assets so the SPA works offline (#693 part 1) #706, fix(web): vendor Swagger UI so /api/docs works offline (#693 part 2) #712,
  closes Web UI and /api/docs depend on cdnjs/jsdelivr — vendor or fallback for offline use #693). DOMPurify, marked, Prism (5 langs + theme), and the
  Swagger UI bundle/css are now vendored under web/static/vendor/
  with pinned SHA-256 in THIRD_PARTY_LICENSES.md. CSP tightens to
  script-src 'self' (no 'unsafe-inline'); a paired positive +
  negative regression guard pins it. Footprint: ~1.7 MB total
  minified.
• Storage: duplicate chunk inserts on shared SQLite DBs (Indexing creates duplicate chunks within a single run (no UNIQUE on content_hash) #691,
  PR fix(storage): block duplicate chunk inserts via UNIQUE + INSERT OR IGNORE (#691) #705) — already in the merged [Unreleased] section. Adds
  UNIQUE(namespace, source_file, content_hash, start_line) +
  INSERT OR IGNORE; migration is automatic on first startup —
  duplicate groups collapse with the highest-access_count + use_count
  row winning (BEGIN IMMEDIATE / COMMIT so two booting processes
  don't both run the cleanup).
• Context overview badge surfaces non-in_sync states (fix(web): surface non-in_sync status in Context overview badge (#692) #704,
  closes Context overview shows '3/3 synced' green badge when missing_target > 0 #692) — the badge now sums missing_target + missing_canonical + out_of_sync + parse_error and displays the
  most actionable status, so multi-runtime divergence stops showing
  as 3/3 synced when half the runtimes are missing targets.
• Home bar chart shows full namespace on hover via a title
  attribute (fix(web): show full namespace on hover in Home bar chart #694).
• **Context: installed_at captured from filesystem max(st_mtime_ns)
  • µs ceiling** (fix(context): installed_at capture from FS mtime + us ceiling (#732) #733, closes Windows: capture installed_at from FS mtime with us ceiling, not wall clock #732) so just-installed assets don't
    false-positive as dirty on Windows under FILETIME / time.time()
    clock-domain skew. POSIX byte-identical for ordinary writes.
• Windows compat sweep (refs Windows: replace POSIX path-separator assertions in tests (/ vs \\) #643, Windows: HOME monkeypatch doesn't reach Path.home() on Windows (uses USERPROFILE) #644) — first batch landed
  this release:
  • mm init no longer crashes on cp1252/cp437 consoles (fix(cli): force UTF-8 stdout/stderr on Windows console (#643 cluster H-1) #728,
    cluster H-1).
  • norm_dir_prefix uses os.sep so memory_dir prefixes match
    chunks.source_file (fix(indexing): use os.sep for memory_dir prefix so Windows paths match (#647) #717, closes Windows: Path canonicalization in indexing — kind detection + memory_dir comparisons #647).
  • source_filter folds path separators across seven MCP tools
    (mem_search, mem_list, mem_consolidate, mem_decay,
    mem_auto_tag, mem_export_chunks, mem_entity_scan) (fix(search): normalize separators in source_filter match (#720) #722,
    closes Windows: search source_filter substring match ignores separator (sibling of #647) #720).
  • categorize_memory_dir handles backslash paths so the mm init
    wizard auto-detects Claude / Codex / Gemini memory folders on
    Windows (fix(config): support Windows backslash paths in categorize_memory_dir (#316) #716, closes Windows: wizard fails to auto-detect Claude/Codex memory dirs (backslash path regex) #316).
  • ~/... path_glob rules expand via .as_posix() so
    pathspec.GitIgnoreSpec actually matches files (fix(config): use as_posix() for tilde-expanded path_glob (#643 cluster E) #726).
  • mm wiki override prints consistent forward-slash separators
    in the Seeded … line so it matches the adjacent git add …
    hint (fix(wiki): use as_posix() for Seeded path display (#643 follow-up) #719).
  • Test-side: set_home helper + ~133 HOME monkeypatch sweep
    (test(windows): set_home helper + sweep ~133 HOME monkeypatch sites (refs #644) #714); POSIX-only skipif markers (test: skipif POSIX-only assertions on Windows (mode bits, fcntl, signal) — Windows CI PR 2/N #713, test(windows): skipif POSIX-only pwd.getpwuid simulation in fastembed_cache (#643) #721, test(windows): skipif POSIX-only claude-projects discovery tests (#643) #725, test(windows): skipif POSIX-only unlink semantics in test_force_overrides_db_lock (#643 cluster H-2) #729);
    path-separator normalisation (test: normalize path separators in 7 cross-platform assertions (#643) #711, test(windows): normalize path separators in 8 wiki override assertions (#643) #718); shutil.which for
    the mm entry-point lookup (test(windows): use shutil.which for mm entry-point lookup (#643) #723); Windows-friendly
    test_init_cmd.py (test(windows): make test_init_cmd.py Windows-compatible (#643) #724).

Documentation

• mm embedding-reset needs an idle DB (docs(embeddings): warn that embedding-reset needs an idle DB #710, closes Embedding-loss race when two processes use different models on the same DB (follow-up from #691) #707) —
  callout in configuration.md#reset-flow and a one-line warning in
  embeddings.md "Switching Models on an Existing Index" cover the
  same-dim model-swap race that PR fix(storage): block duplicate chunk inserts via UNIQUE + INSERT OR IGNORE (#691) #705's INSERT OR IGNORE path
  cannot catch.

Release mechanics

This is a behavior-change release (storage migration, install-time
installed_at capture change, CSP tightening). Per
feedback_prerelease_dry_run.md / project_release_workflow.md:

1. After this release PR's CI is green, squash-merge.
2. Push test-v0.1.35a1 tag from the merge sha → TestPyPI dry-run.
3. Approve the pypi environment for the test tag, smoke-test the
   wheel in an isolated worktree (fresh uv tool install), then
   push v0.1.35 from the same sha.
4. Approve pypi again if pending_deployments is non-empty
   (default expectation per the v0.1.31 / v0.1.34 datapoints — the
   v0.1.30 same-sha auto-skip was the exception).
5. gh release create v0.1.35 with notes extracted from this
   CHANGELOG section.

Test plan

• uv run pytest packages/memtomem/tests/test_docs_guards.py -q
  — 11 passed (CHANGELOG schema).
• CI required checks: lint, typecheck, test, notebooks,
  golden-path, CLAAssistant.
• TestPyPI dry-run: fresh uv tool install memtomem==0.1.35
  from https://test.pypi.org/simple/, confirm the storage
  migration log line emits exactly once and mm --version
  prints 0.1.35.
• Production: uvx --refresh --from memtomem==0.1.35 mm --version
  polls clean.

🤖 Generated with Claude Code