channel-memory: add async LLM digest worker and cost caps

[mostlydev]

Motivation

Deterministic channel-memory digesting is the safety baseline, but the Phase 2 design also needs an off-hot-path LLM worker for higher-quality topic and sequence rollups. This must not add latency to channel-awareness serving or blur exact source provenance.

Depends on #265. It can land after #267 if deterministic raw_window+digest ships first.

Scope

• Add an async worker that processes queued source-message windows into structured digest blocks.
• Keep hard events verbatim or near-verbatim; use lossy prose only for sparse topic_rollup and sequence_rollup style blocks.
• Require structured JSON output and reject malformed or provenance-free results.
• Cache work by source message ids plus content hashes.
• Store model/provider/version, generated time, sparse flag, source window, and source-message provenance for every generated block.
• Enforce conservative per-channel/per-pod daily call and cost caps with clear telemetry.
• Fall back to deterministic-only digest output when the worker is disabled, over budget, or failing.

Out of scope

• The base SQLite adapter schema beyond extensions required for queued work and processor metadata.
• claw-wall push ingest and channel-awareness serving changes.
• Public YAML tuning surfaces beyond internal environment overrides.
• Release pin, changelog, or nav changes.

Acceptance

• Unit tests cover queue ordering, malformed LLM output rejection, cache reuse by content hash, cost-cap enforcement, and deterministic fallback.
• Processor tests prove edit/delete/forget marks affected derived blocks stale or dirty.
• Fixture tests show older verbose channel material compresses into sparse blocks while hard events remain faithful.
• No LLM call occurs during /digest hot-path lookup.
• Public docs and examples stay deployment-agnostic.

Refs #262
Refs #263