evaluation: validate multi-resolution context projections and progressive retrieval

Feature description

Define multi-resolution context projections and progressive retrieval under a fixed PreparedContext budget.

The design should preserve immutable full content as authority while using rebuildable summaries or abstracts to select and initially inject more relevant items, expanding to exact full content only when needed.

Problem and proposed solution

Current PreparedContext selection is bounded and citation-aware, but long Memory and Experience content competes directly for the same byte budget. #1356 uses "progressive disclosure" for explaining a prepared selection; this issue concerns retrieval and injection behavior, not diagnostics.

The RFC should evaluate a three-level projection model:

• L0: immutable authoritative full content referenced by the existing exact citation;
• L1: rebuildable overview suitable for bounded PreparedContext inclusion;
• L2: short rebuildable abstract or embedding text suitable for candidate selection.

Required design points:

• projections include the authoritative content digest, generation profile/version, and projection status;
• projection generation never blocks or changes the authoritative write;
• no-model, generation-failure, stale, or invalid projection paths fall back safely to L0;
• exact-term FTS behavior remains available and is not silently replaced by summary-only search;
• Runtime owns layer selection, expansion, deduplication, citation, and Unicode-safe byte budgeting;
• exact L0 reads continue through existing versioned get/citation behavior rather than a synonymous unversioned expand API;
• scope_id remains an opaque business partition and is not parsed as a hierarchy;
• enablement remains experimental until a frozen evaluation demonstrates value.

Acceptance criteria

• Every L1/L2 projection maps to one exact L0 identity and is invalidated/rebuilt when its content or generation profile changes.
• Missing, stale, invalid, or unavailable projections cannot make authoritative content unreadable.
• PreparedContext citations always identify the exact authoritative Revision, never only a summary row.
• Exact-term regression cases remain retrievable even when the term is omitted from L1/L2.
• Evaluation reports answer success, citation correctness, injected bytes/tokens, items accommodated, summary omission, exact-L0 expansion rate, latency, inference cost, and abstention.
• The progressive policy is compared against the existing baseline on long-context cases from the shared evaluation infrastructure.
• Diagnostic Receipts, when available from #1356, report selected level, fallback, omission, and expansion reasons without retaining content.
• The public prepare contract remains compatible unless a separately accepted API change is required.

Alternatives considered

• Summarize and overwrite authoritative Memory: loses evidence, detail, and exact revision semantics.
• Use only L2 for full-text and vector search: can miss identifiers and exact terms that summaries omit.
• Add a new unversioned expand endpoint: duplicates exact citation reads and risks resolving the wrong Revision.
• Treat Project/Workstream scopes as a retrieval tree: scope organization and content resolution are independent problems.

Additional context

Related: #1356, #1263, #1359, and the current Memory search/reranking RFC.

• I am willing to contribute code, docs, or design feedback.