[Feature] Semantic compression — rolling summaries and project digests

[rahilp]

Why

Per-query synthesis already ships in recall (v1.4.0, issue #59) — that solved the "dumb retrieval" problem for single queries. What's still missing is the write-back loop: when a tag accumulates dozens of raw entries, there's no compression happening. Users returning to a heavily-used tag get 15 raw memories instead of one clean synthesis. The retrieval quality problem is solved ephemerally; the storage bloat problem is not.

What this covers

• Rolling summaries: a nightly cron job identifies tags with >20 entries and compresses related clusters into a single synthesized entry, preserving source IDs in metadata as provenance links. Originals are not deleted.
• /digest?tag=X endpoint: on-demand synthesis of everything stored under a given tag, returning a structured "state of the world" paragraph rather than raw entries. Useful for project handoffs, onboarding a new AI session, or sharing project context.
• synthesized tag convention: summary entries are tagged synthesized and linked back to source IDs so callers can distinguish generated summaries from raw memories.

Out of scope (already shipped)

Per-query synthesis on recall responses (issue #59, v1.4.0). This issue is about writing synthesized entries back to storage, not enriching retrieval responses.

Implementation notes

• Compression can reuse the /chat endpoint pattern (Workers AI + Llama already in place)
• Cron trigger fires nightly; only processes tags with >20 entries to avoid unnecessary runs
• deleteByIds is not called on originals — summary is additive
• Blocks [Feature] Memory consolidation — clustered summaries with provenance links #15 (memory consolidation) and #18 (agent-optimized context packaging)