Add memory provider

[fivetran-davefowler]

Summary

Adds a tool-agnostic memory provider: durable, agent-facing notes anchored to the warehouse objects where they're relevant. Supersedes #11 (that PR's branch notes_provider was renamed to memory-provider, which closed it).

• agents-schema memory --memory-file memory.yml → writes AGENTS.MEMORY + AGENTS.MEMORY_ANCHOR, registers the memory provider in AGENTS.ROOT
• Adds the matching reusable workflow + composite action and memory-setup.md, mirroring dbt/Looker/OSI

Positioning (what changed from #11)

Memory is framed as the lightweight path for teams without a semantic layer. If you run OSI, object-local ai_context already carries these notes and memory is largely redundant — the SPEC and setup guide say so plainly rather than overselling. Anchors target physical objects (schema/table/column) on purpose: a column anchor is reachable from raw schema and from an OSI-aware consumer (OSI fields map down to columns); the reverse isn't true. metric and relationship are the only logical anchors, for things with no single physical column.

Schema / design decisions

• confidence is FLOAT in [0,1] (was free-form varchar) so consumers can threshold and rank.
• Relationship anchors carry the join inline via from_*/to_* with positionally-paired from_columns/to_columns arrays (handles composite/>2-column keys, mirrors OSI_RELATIONSHIP). relationship_name is an optional, unenforced pointer to OSI_RELATIONSHIP.name when OSI is present — memory does not depend on OSI.
• Provenance stays flat on MEMORY (source + confidence) for v1; a separate MEMORY_SOURCE edge table is deferred.
• Tighter anchor validation: rejects locator columns that don't belong to the anchor's type, plus the existing fail-fast checks (unknown fields, bad scalar types, dup memories/anchors, invalid anchor types, missing locators) and confidence range.
• Key naming follows the OSI parent/child convention (MEMORY.memory_id ↔ MEMORY_ANCHOR.memory_id), documented in code and SPEC.

Verification

uv run python -m unittest discover -s tests
Ran 21 tests
OK

git diff --check   # clean

🤖 Generated with Claude Code