[ScriptedAlchemy]

Problem

Both observation hydration adapters awaited one SQL round trip for each distinct canonical body. The temporal projector already batches observation rows, but then reintroduced a per-body waterfall while decoding each row. Concurrent tasks on the same retained snapshot would not remove that dispatch overhead.

Change

• Read one observation's distinct canonical-body references in one indexed SQL statement, shared by the global observation decoder and temporal-store decoder.
• Keep the caller's existing read connection/snapshot. No spawned tasks, additional reader owners, cache, dependency, schema change, or raised budget.
• Preserve the zero-query inline path. A left join retains missing references and input ordering; every present body still passes the existing encoding, declared-size, and SHA-256 verification before canonical observation reconstruction.
• Keep the existing exact-SQL materialization ceiling and surrounding bounded observation paging. This batches one observation, not an entire store or unbounded request cohort.

Regression coverage

Added a behavioral regression in each adapter using the actual SQLite-backed TestConnection. A query-counting wrapper forwards every call to the real reader. Tests cover two distinct bodies plus a shared duplicate, exact reconstruction, one body query instead of two, no query for inline content, and typed missing/corrupt-body refusals. The global adapter also checks that an unrelated corrupt body is not fetched. Existing tests remain intact.

Validation actually performed

Passed six native SQLite 3.46.1 mechanism checks using the exact SQL extracted from the changed source: 32 referenced bodies required one statement rather than 32, identity/deflate/Unicode rows matched scalar reads, missing rows and ordering were preserved, unrelated rows were excluded, 100 randomized reference subsets matched, and query-only execution left data unchanged. EXPLAIN QUERY PLAN uses the canonical-body primary-key index for each requested hash. The supplementary Python verifier exercised corruption/declared-size cases; it does not validate the Rust implementation.

The local source copies were verified against their original Git blob hashes before editing. git diff --check passed. Published blobs match the reviewed local files. A separate source-review pass checked snapshot ownership, error mapping/order, integrity checks and both production callers. Publication is based on refreshed master adbf0e39; its intervening Doctor merge does not touch these files.

Not run: Rust compilation, Rust tests, rustfmt, Bazel, production latency/RSS benchmarks, or the full native matrix. This environment has no supplied full checkout or Rust/Bazel toolchain. Keep this draft until hosted validation executes the owning global-db, session-temporal-store and store unit tests. No wall-clock speedup is claimed from the statement-count reduction.

This is one focused result of the requested async hot-path audit. It does not replace the independent retention/GC work in #4718 or modify its branch. Existing PR repair/merge automation owns CI follow-through and authorized merging.

[ScriptedAlchemy]

Consolidated into #4742 (integrate/open-pull-requests) at the captured head shown in its description, as requested. Continue review and further fixes on #4742. This closure does not assert CI readiness; the replacement remains draft. Source branch retained.