wasteNot-adjacent: hub blackboard kind (independent-write brainstorm alternative)

New brainstorm kind: opt-in, independent-write coordination, replacing DM fan-out for the questions that don't need reaction. Two live trials this session validated it (coder02-terra-cc, 4 lanes/filesystem; pm-llmmsgsrv-cc, 4-agent panel on real Postgres-role questions for Elazar) — this epic captures what those trials proved and specs the hub-native build. SPEC, derived from the trials: 1. Storage: v2.sqlite table (new, e.g. blackboard_entries + blackboard_responses), not a filesystem dir. Server-set write timestamp on each response row — this is the fix for the one real weakness the first trial surfaced: file-per-agent self-reported ordering is unverifiable after the fact ('I wrote before reading' is agent-honesty, not provable). A DB row with a server clock is. 2. Additive, not a replacement: existing brainstorm (DM fan-out, close_quorum) stays for reactive cases — agents genuinely needing to see and respond to each other. Blackboard is for independent-judgment questions only. Both trials were explicitly non-reactive; neither trial tested the reactive case, so don't assume this generalizes there. 3. Opener creates an entry, names the expected respondent set, hub does ONE notify per respondent (batched respecting ARRAY_BLAST_MAX=4 — a panel over 4 needs multiple notify batches, same as any other multi-recipient send). 4. Respondents write once. Hub does not expose other respondents' answers to a respondent who has not yet written their own — enforce independence structurally at read-time, not just by instruction (the second trial only had voluntary compliance; a real adversarial case wouldn't). 5. Completion: hub closes the entry and notifies the OPENER exactly ONCE, when the full expected set has written — same close_quorum shape existing brainstorm already has. CRITICAL DESIGN CONSTRAINT, learned live this session by getting corrected by Elazar mid-trial: the hub must NOT push a per-response event to the opener as each respondent writes. That recreates the exact N-multiplier cost blackboard exists to remove, just moved one hop downstream (agent-to-agent chatter becomes opener-to-human chatter). Only the single completion (or timeout) event reaches the opener. 6. Timeout/escalation, same shape as the existing brainstorm-never-hang rule: an entry not complete after N minutes (default 4, Elazar-set this session) triggers ONE straggler DM from the hub (or the opener, TBD which layer) asking status, not a broadcast. If that gets no reply, fall to the fleet's agent-liveness diagnostic (ask bin-venus-cc for the current canonical script/path — do not hardcode a name that may have moved) to distinguish idle/compacting/wedged/dead before treating it as a real block. 7. Real validation from trial 1: forced independence produced a genuine find (a lane read the actual INSERT function and found zero array validation, reframing the whole question) that a DM thread's first-responder framing would likely have suppressed. Worth keeping as the design's stated rationale, not just the cost argument. Trial artifacts: /data-venus/terra/brainstorm/2026-09-02-terra136/ (trial 1), /data-venus/llmmsg-srv/blackboard/2026-09-02-pgroles-q1q2/ (trial 2, real Postgres-role decision for Elazar, 4/4 responded, unanimous, useful result — proof this works on live decision-relevant questions, not just synthetic ones).