refactor(core): simplify v0.23 runtime protocols

[phernandez]

Why

The v0.23 indexing and semantic-search work added a large Protocol surface so the
portable runtime could serve both Basic Memory and Basic Memory Cloud. A
cross-repository audit found a smaller set of contracts that were unused,
duplicated, mutable despite representing results, or hiding ownership behind an
extra adapter.

This cleanup removes those vestiges before the release without removing any
feature or weakening a capability that Cloud consumes.

What Changed

• Removed unused or duplicate Protocols for directory deletion, initial project
  indexing, relation scheduling, note-file cleanup, and vector synchronization.
• Added one shared, frozen VectorSyncBatchResult and kept mutable batch
  accounting private to the semantic-vector repository operation.
• Made structural data members read-only so frozen dataclasses and ORM-backed
  values can satisfy the same narrow contracts.
• Removed unnecessary nominal Protocol inheritance from concrete embedding,
  reranking, SQLite, pgvector, and Milvus implementations.
• Moved local missing-file race handling into the local storage metadata adapter
  and removed the wrapper that previously obscured that responsibility.
• Made local frontmatter result values frozen and slotted.
• Ignored all repository-local .codex runtime state and nested worktrees while
  allowlisting a minimal shared .codex/basic-memory.json.

There are no user-visible feature removals.

Implementation Details

src/basic_memory/runtime/vector_sync.py is now the single portable home for
the vector-sync capability and its immutable batch result. The repository
implementation uses a private accumulator and freezes counts, timings, and
failed entity IDs at the public boundary.

Property-only Protocol members are now read-only properties. Concrete providers
remain structurally compatible, but no longer inherit those Protocols
nominally; this avoids inherited property descriptors conflicting with normal
instance initialization.

LocalStorageFileMetadataSource now owns checksum lookup semantics directly. A
wrapped FileNotFoundError still means the file vanished between existence and
checksum checks, while permission and transient I/O errors continue to
propagate.

The shared Codex config contains only the coding profile and canonical
repository identifier. User project mappings and all .codex/worktrees state
remain machine-local.

Testing

Automated

• just fast-check: passed (Ruff fix/format and ty type checking; nine
  existing Python 3.14 event-loop deprecation warnings).
• BASIC_MEMORY_ENV=test LOGFIRE_IGNORE_NO_CONFIG=1 uv run pytest -q --no-cov tests/index/test_local_schedulers.py tests/index/test_inline_storage_event_processor.py tests/indexing/test_embedding_index_planning.py tests/indexing/test_vector_sync_planning.py tests/repository/test_semantic_vector_sync.py:
  51 passed.
• Expanded provider/vector regression suite: 294 passed.
• just doctor: passed the isolated file, database, indexing, vector, search,
  and status flow.
• just fast-test: 4,969 passed and 33 skipped. The only failure was the live
  postgres-openai semantic-quality benchmark, which reached OpenAI and
  received HTTP 429 insufficient_quota.
• AST Protocol audit: no writable Protocol attributes remain.
• Retired-contract name scan and git diff --check: passed.

Risks / Follow-ups

• The live OpenAI benchmark remains dependent on account quota; its failure did
  not exercise a changed local code path.
• A broad SearchRepository split and complete cross-repository frontmatter
  result unification are intentionally deferred until after v0.23. They would
  expand release risk without simplifying this patch.