…r (Goal 7)

Final RFC §CLI/Web UI surface — the temporal-validity window is now
visible end-to-end. Backend exposure on ``ChunkOut`` (used by every
chunk-shaped route response: list, get, search results, similar)
threads ``valid_from_unix`` / ``valid_to_unix`` from chunk metadata.
Frontend renders a compact ``Valid: 2025-08-15 → 2026-03-31`` badge
next to the existing created-time badge in the detail modal and inside
each search result item. Always-valid chunks (both bounds null) hide
the badge entirely so the default UI stays compact for users who
haven't opted into temporal-validity frontmatter.

Expired chunks (``now > valid_to``) get a muted opacity + line-through
on the badge text, with a localized "Expired" tooltip. The chunk stays
visible — validity is a search-time filter concept (Goal 4 stage), not
a storage-level hide. The Web UI is a viewer; filtering happens
upstream when the user issues a search with ``as_of``.

Date display uses ``YYYY-MM-DD`` regardless of the original frontmatter
shape (date vs quarter), since the unix-second column does not preserve
that distinction. ``∞`` renders for unbounded sides — half-bounded
windows (only ``valid_from`` or only ``valid_to`` set) are first-class.

Drive-by fix in ``PATCH /chunks/{id}/tags``: the route reconstructed
``ChunkMetadata`` with an explicit eight-field list, silently dropping
any field not enumerated. Goal 7 adds two more such fields, so the
implicit-loss bug would have ridden along with this PR — switched to
the dict-spread copy-with-override pattern that the ``mm add`` CLI
already uses (memory.py). This also implicitly restores
``overlap_before/after``, ``parent_context``, ``file_context``, and
``language`` round-tripping through tag updates — those have been
silently lost since the route was written. Test pins the regression so
re-flattening is loud.

i18n: two new keys (``search.detail_validity_label``,
``search.detail_validity_expired``) added to en.json + ko.json
("Valid"/"유효", "Expired"/"만료됨") so the trust-UX surface stays
localized. Date format itself is locale-independent (ISO-8601 strict).

Tests (3 new, in test_web_routes.py at TestChunkValidityFields):

- ``test_chunkout_includes_validity_when_set``: chunk with both bounds
  populated (2024-12-15 → 2025-Q1 end) round-trips through
  ``GET /api/chunks/{id}`` with the unix integers intact.
- ``test_chunkout_validity_null_when_unset``: default ``_make_test_chunk``
  produces a chunk without validity → both fields serialize as ``null``
  so the frontend always-valid branch (hidden badge) fires.
- ``test_tag_update_preserves_validity``: regression for the
  drive-by fix — PATCH /chunks/{id}/tags must not drop validity (or
  ``parent_context`` / ``overlap_before`` for that matter). Inspects
  the ``upsert_chunks`` await args directly because that is the wire
  the silently-dropped fields would have been lost on.

Frontend rendering coverage is manual (no JS test harness in this repo);
the API contract test pins what the frontend reads from JSON.

This is the last code Goal in the RFC — no further surface work
planned. Goal-by-goal status table updated in the matching
memtomem-docs commit.

Co-Authored-By: Claude <[email protected]>