…-from tags

PR-1 of the chunk_links series (private RFC
`mem-agent-share-chunk-links-rfc.md`). Storage-only change — no public
API surface yet, no behavior change for existing callers.

`mem_agent_share` historically encoded provenance as a
`shared-from=<uuid>` audit tag baked into the destination chunk's
`tags` array. Tag-based provenance has three problems: it doesn't
benefit from an index (fanout query is `tags LIKE '%shared-from=%'`,
a full table scan), it breaks on UUID churn (reindex re-issues chunk
ids and the audit chain breaks at the gap), and it can't enforce a
relationship across delete.

This PR adds the storage substrate. PR-2 wires the writer
(`mem_agent_share` records into `chunk_links` on share) and reader
Python API; PR-3 (optional) exposes a `mem_share_lineage` MCP tool.

## Schema

`chunk_links` has `PRIMARY KEY (target_id, link_type)` so every
destination chunk has at most one link of each type, plus indexes on
`(source_id, link_type)` and `namespace_target` for fanout / per-NS
audit queries. FK semantics:

- `ON DELETE SET NULL` on `source_id` — preserves the existing
  copy-on-share durability: a teammate deleting their note does not
  delete yours; the link row stays with `source_id=NULL` and the
  destination chunk lives on. Provenance is still recoverable from
  the markdown `shared-from=` tag (still written into content).
- `ON DELETE CASCADE` on `target_id` — destination delete drops the
  row, no dangling pointer.

`link_type` validation lives in Python (`_VALID_LINK_TYPES`) rather
than as a CHECK constraint so adding a new value (`consolidated_from`,
`reflected_from`) is one PR, not two.

## Back-fill

Existing share copies (created before this PR ships) have a
`shared-from=<uuid>` tag in `chunks.tags` but no row in `chunk_links`.
A one-shot pass scans those rows, parses the source UUID, resolves it
against `chunks.id` (NULL if the source was already deleted), and
`INSERT OR IGNORE`s into `chunk_links`. Completion is recorded in
`_memtomem_meta` (`chunk_links_backfill_v1`) so subsequent startups
short-circuit. Bumping the version key triggers a re-run if the
parser ever needs to widen.

## Tests

- `test_chunk_links_schema.py`: table+indexes shape, idempotent re-run,
  FK SET NULL on source delete, FK CASCADE on target delete, PK
  uniqueness conflict.
- Back-fill cases: existing source resolved, missing source → NULL,
  unrelated tags ignored, marker prevents re-scan of post-migration
  rows, malformed `tags` JSON skipped.

Full suite: 2421 passed (was 2411, +10). ruff clean.

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