[engine] carry typed-action-contract schemas on the Sources payload

[cmttt]

Problem

#55 added the typed-action-contract graph specializer + a schema loader that reads schemas from a filesystem directory (OSPREY_SCHEMAS_DIR, else OSPREY_RULES_PATH/schemas). A worker whose rules come from etcd has no schemas directory on disk, so resolve_schemas_dir() returns None and the specializer never activates there — even with the prune/shadow allowlist set.

Approach

Carry the schema JSON (schemas/*.json + schemas/types/*.json) on the same Sources payload as the rules, so it rides the existing rules→etcd→worker rail, and have the loader read schemas from the in-memory Sources. The serialized payload (to_dict) is byte-identical to today when there are no schemas, so the publisher/provider/deployer are untouched.

Changes

• Sources — a separate _schemas map (kept out of the .sml compile path): from_path (anchored glob('schemas/**/*.json')), to_dict (merge; no-op when empty), from_dict (partition schemas/ keys before the .sml-asserting add_source), hash() (fold schemas in so a schema-only edit invalidates dedup and the worker reloads), and schemas()/get_schema() accessors.
• schema_loader — extract a source-agnostic parse_schema(raw, ref_reader, where); load_schema stays the disk wrapper (behavior unchanged); add load_schema_for_action_from_sources(action, schemas_map) with a key-space $ref traversal guard.
• typed_contract_dispatch — load_and_register_specialized_graphs(..., schemas=None) reads from the in-memory map when present, else falls back to the on-disk dir (local-dev preserved).
• both engines — pass self._execution_graph.validated_sources.sources.schemas().

Rollout (worker-before-deployer, one-way)

The payload format gains schemas/ keys, which an old worker's from_dict would route into the .sml-asserting path. So the worker must run this code before any payload carrying schema keys is published. Until schemas are pushed, to_dict is byte-identical to today, so the change is fully back-compatible during the rollout window.

Tests

72 passing: Sources round-trip + back-compat (no-schema payloads, schema-key payloads, hash-on-schema-edit), schema_loader disk/Sources parity + $ref traversal guards, dispatch source-selection (Sources vs disk vs disabled). Existing disk-loader tests unchanged.

[cmttt]

Verified end-to-end locally with osprey-coordinator-asyncio + osprey-worker-asyncio running (schemas sourced from the in-memory etcd Sources, no OSPREY_SCHEMAS_DIR):

Loaded 2 specialized graphs from Sources (prune=['guild_invite_created'] shadow=['guild_invite_created','guild_member_kicked'])
TAC-PROOF: action=guild_invite_created dispatch=PRUNE(serve specialized graph) registered_specialized=True
TAC-PROOF: action=guild_member_kicked dispatch=SHADOW(serve full + diff) registered_specialized=True

The first live run surfaced a real gap: the async sink (AsyncRulesRunner.classify_one) bypassed engine.execute() (called async_execute() on the full graph directly), so the #55 dispatch was a runtime no-op on the async worker. Fixed in cd7ccb0 (+ regression test). After the fix, both actions hit the specialized path through the live coordinator→worker bidi transport.