Fix AtomicMemory thread scope and add meta-fact filter

[ethanj]

Summary

Fix the AtomicMemory provider dropping thread scope (session ID) on memory operations, and introduce a MetaFactFilter that removes extraction artifacts — facts the extractor emits about the conversation itself rather than about the user — before they can pollute embedding search results.

Changes

• Thread scope fix — scope-mapper.ts and handle-impl.ts now correctly carry thread/session scope through list, search, and upsert paths so memories are isolated to the right conversation thread.
• MetaFactFilter (src/memory/meta-fact-filter.ts) — pattern-based filter that rejects statements matching extraction artifacts (The user asked for…, As of <date>, X is a term mentioned…, A name was mentioned…). Exported from src/memory/index.ts.
• Provider types (types.ts) — add threadId / session scope fields to internal request/response shapes.
• Mapper updates (mappers.ts) — propagate new scope fields through wire-format mapping.
• Fixture refresh — search.raw.json, search-fast.raw.json, list.mapped.json, and related mapped fixtures updated to reflect the new scope fields in provider responses.
• Test coverage — atomicmemory-provider.test.ts extended for thread scope round-trips; meta-fact-filter.test.ts added with pattern coverage across all rejection categories; namespace-base-routes.test.ts added for scoped route construction.
• AlignBench benchmark suite (benchmarks/alignbench/) — controlled 60-query / 55-fact recall benchmark with a standalone runner, six-model embedding ablation, and full results. Validates the meta-fact filter lift and falsifies the pronoun-rewrite hypothesis against a pre-registered threshold.

Why

Thread scope was being silently dropped in the scope mapper, causing all memory reads and writes to fall through to the namespace root rather than the per-session partition. This meant thread-isolated memories were readable across unrelated sessions.

Separately, production extraction pipelines emit low-signal meta-facts (The user asked for the user's name., As of May 14, the user is a term mentioned in the conversation.) that occupy the same embedding neighborhood as real user facts, causing a retrieval margin collapse — the correct fact loses top-1 to an extraction artifact. The AlignBench benchmark quantified this: dropping extraction meta-facts yields the largest single recall lift (+0.03–0.05 r@1) of any algorithmic retrieval patch tested.

Validation

• pnpm test — all provider, meta-fact-filter, and namespace-base-routes tests pass.
• pnpm typecheck — no new type errors; new threadId fields are fully typed.
• Fixture contract tests confirm the updated wire shapes match the refreshed recorded responses.
• AlignBench run results committed in benchmarks/alignbench/runs/ for reproducibility; the clean-pool variant (meta-facts removed) shows r@5 0.950 vs 0.933 baseline with distractor-top1 dropping to zero.