MC-4365

Build the personalized NBLM/radio-show curation layer discussed with Elmar. Goal: Create the upstream editorial pipeline for the daily NotebookLM-style audio overview so the show is personalized, recent, and not repetitive. Context / user requirements: - Keep NotebookLM audio generation as the studio/output layer. - Improve story discovery and selection before NotebookLM. - Use Elmar's active context and interests: Hermes/Mission Control/agent orchestration, Claude Code/Codex/Hermes memory, AI trends, aviation/FlySafair, South African business/regulatory news, peptides/longevity, geopolitics/energy when business-relevant. - Add a dedicated Boeing/Airbus narrowbody fleet watch relevant to a South African low-cost airline: 737/A320neo delivery delays, CFM LEAP and Pratt & Whitney GTF supply/fixes, spare engines, MRO/shop bottlenecks, supplier/logistics pipeline, lease-rate/capacity economics, production caps/quality holds. - Avoid stale articles leaking into NotebookLM. Default to last 72h; aviation/fleet watch up to 7 days; older sources only if explicitly marked background and not presented as news. - Avoid repeating daily stories like Spirit Airlines unless there is a material update. Hormuz-like recurring topics may appear only as short material-update/watchlist items. Deliverable: 1. Inspect the existing NBLM/audio/research pipeline, scripts, tasks, and relevant skills. 2. Design and implement a story-ledger/deduper and freshness filter. 3. Implement an Elmar Interest Map / active-context topic generator using available local sources where safe: session history, MC tickets/reports, markdown plans/notes, durable preferences. 4. Implement candidate collection lanes and scoring/selection rules. 5. Add or update the daily radio/NBLM task so it creates a curated NotebookLM source pack/slate, with explicit labels TODAY / THIS WEEK / BACKGROUND / EXCLUDE. 6. Add tests/dry-run output showing: selected slate, suppressed duplicates/stale stories, source ages, rationale, and NotebookLM prompt/source-pack text. 7. Document the runbook and config knobs. Acceptance criteria: - A dry run can produce a personalized story slate without generating audio. - Dry run shows at least: story title, source URL, source age, topic lane, novelty/material-update decision, relevance rationale, and whether it is included/suppressed. - Duplicate stories from the last 14-30 days are suppressed unless material_update=true. - Articles older than the configured freshness window are excluded or marked BACKGROUND and cannot be narrated as fresh news. - Boeing/Airbus narrowbody fleet-watch lane exists and is weighted for SA low-cost airline relevance. - NotebookLM source pack is clean and instructs NotebookLM not to rehash stale background. - No production posting/Telegram/audio generation occurs during tests unless explicitly enabled. - Expected check-in: today after discovery/design pass, before risky integration changes. Implementation notes: - Follow the single-session dev-loop for source edits. - Prefer deterministic collectors/scoring; use LLM only for clustering/material-update/editorial judgment where it adds value. - If GLM/runtime skill pickup is unreliable, encode the workflow explicitly in scripts/runbooks rather than relying on a model remembering to invoke a skill.