Turn async accessor methods to sync properties for provider attributes

[gvanrossum-ms]

Plus some changes to AGENTS.md -- I want to start using worktrees (so we can do multiple projects in parallel), but the agent got confused and was looking in all the wrong places (so we added more hints to AGENTS.md).

[gvanrossum-ms]

@KRRT7 @bmerkle

[KRRT7]

Hey Guido, would you mind splitting the AGENTS.md changes into a separate PR? Keeps the sync properties refactor easy to review on its own.

[KRRT7]

Sync properties refactor looks good — none of the old async getters were doing actual async work, so @property is more honest and Pythonic.

Two things worth noting in the PR description:

1. The sqlite provider's old get_conversation_threads() was creating a new ConversationThreads instance on every call (with a lazy import). Now it creates one in __init__ and returns the same instance. That's a behavioral change / bug fix — you'd lose thread state between calls before.

2. term_to_semantic_ref_index was renamed to semantic_ref_index on the sqlite provider to match the interface. Minor but worth calling out.

[gvanrossum]

Hey Guido, would you mind splitting the AGENTS.md changes into a separate PR? Keeps the sync properties refactor easy to review on its own.

Just ignore AGENTS.md.

1. The sqlite provider's old get_conversation_threads() was creating a new ConversationThreads instance on every call (with a lazy import). Now it creates one in __init__ and returns the same instance. That's a behavioral change / bug fix — you'd lose thread state between calls before.

It's a bugfix, but it doesn't actually change behavior -- there's only one non-test call site, in ConversationSecondaryIndexes.create(), which sets self.threads and then that attribute is used elsewhere. (It's a bug, because the memory provider doesn't create a new one -- and that provider is the original one.)

On top of that, I don't think the actual ingest and query algorithms even uses the threads attribute. At least the new version is consistent. :-)

2. `term_to_semantic_ref_index` was renamed to `semantic_ref_index` on the sqlite provider to match the interface. Minor but worth calling out.

It's a bit awkward: the class is TermToSemanticRefIndex and it is stored on the class as private _term_to_semantic_ref_index, but the callers (almost?) all store it into a variable named semantic_ref_index (or index, in at least one case). The class name covers what it does well (it maps terms to semantic ref IDs) but it's too awkward, so everyone shortens it -- even the sqlite schema uses SemanticRefIndex for the table name.

I don't believe any action is necessary, but I'll wait for approval from @bmerkle so I don't have to cheat to merge it myself.

[bmerkle]

hey, great work. LGTM. no additional feedback from my side.
had to merge in main, so @gvanrossum has to merge it