Add example notes_on provider

[fivetran-davefowler]

Summary

Adds notes_on, an optional Agents Schema provider under examples/providers/notes_on for teams that want a durable note layer on warehouse schemas, tables, and columns. It is useful as-is for agents, catalogs, review workflows, or analytics tooling that need object-scoped notes.

This is not a built-in/core provider and does not require everyone adopting Agents Schema to adopt this note shape. It does not change the core agents-schema CLI, core ROOT_ENTRIES, reusable workflows, package metadata, or the official SPEC.md.

The provider publishes durable object-scoped notes using provider-prefixed tables:

• AGENTS.NOTES_ON_SCHEMATA
• AGENTS.NOTES_ON_TABLES
• AGENTS.NOTES_ON_COLUMNS

It also upserts provider-authored AGENTS.ROOT rows at runtime. Those rows are deployment/provider metadata, not new core registry entries.

Implementation

The package includes:

• a standalone agents-schema-notes-on CLI
• YAML source parsing and validation
• Snowflake publishing for the three NOTES_ON_* tables
• provider-owned AGENTS.ROOT upsert behavior
• sample YAML input
• standalone SQL snippets for tables, root rows, and context aggregations
• unit tests for parsing, validation, SQL binding, credentials, CLI errors, and publish flow

Positioning

This reframes the old memory idea as an optional provider rather than a core Agents Schema feature. The shape is intentionally narrow: physical schema/table/column notes, no agent memory state, no write-back semantics, no polymorphic anchors, and no requirement that other providers or consumers adopt these opinions.

The useful distinction from database comments is still preserved: multiple authors, provenance, tags, confidence, importance, and history without overwriting a single canonical object comment.

Context view compatibility

The provider id is notes_on, so the table names line up with the context-view naming scheme from #12. Since the source tables are one row per note, any AGENTS.SCHEMATA/TABLES/COLUMNS enrichment should aggregate to object grain before joining. The example uses a notes_on_notes VARIANT array rather than concatenated text so note boundaries and provenance are preserved.

Verification

uv run python -m unittest discover -s tests
Ran 7 tests
OK

PYTHONPATH=examples/providers/notes_on/src uv run python -m unittest discover -s examples/providers/notes_on/tests
Ran 8 tests
OK

uv run python -m compileall -q src examples/providers/notes_on/src
git diff --check
clean