docs: inventory papers, protocol pins, and reproduction plan

[sdntsng]

Summary

Bounded inventory only. This PR adds one research document that maps the three paper tracks (OpMem, vertical packs, conviction/stop-loss), pins the current Brainy benchmark protocol, maps every live claim to code/data/artifact/SHA, explains why drafts #185 / #186 / #187 stay open, and lists the smallest next experiments with call/cost/runtime estimates.

No benchmark was run. No provider credits were spent. No scores, README percents, or cycle-closeout numbers changed. Anti-benchmax and the independently verifiable 90% LoCoMo / 80% LME / 13/13 / 17/17 goal are preserved as gates, not claims.

Changes

• docs/research/inventory-reproduction-plan-2026-09-18.md — the inventory (papers, protocol, claim map, PR relationship, costed next experiments, reproduce commands marked do-not-run).
• docs/research/README.md — index row (labeled not a pin).
• docs/research/paper-topics.md — pointer to the inventory.

Test plan

• No Go / eval / provider commands in this pass (explicitly out of scope).
• Document is based on origin/dev @ bbe55f7 plus read-only GitHub state of docs: n=1540 product-recall handoff for next-cycle planning #185/Eval-protocol-v2 staged runner and worker slot keep-alive #186/S5: versioned RRF retrieval for search and /recall #187.
• Reviewer: confirm Paper 1–3 paths and “not ready” verdicts.
• Reviewer: confirm E0 (OpMem/marketing on bbe55f7) then E2 (new-tenant S0 180) as the next measurement sequence — or reject.
• Do not treat this PR as a LoCoMo pin or a merge of docs: n=1540 product-recall handoff for next-cycle planning #185–S5: versioned RRF retrieval for search and /recall #187.

Linear / tracking

Paper roadmap: paper-topics.md. Related Linear still open: ENG-76 (finance pack), ENG-86 (supersession; API exists, OpMem still uses /correct), ENG-78 (finance goldens). No new Linear id.

Risk / follow-ups

• Inventory SHA is bbe55f7. Public 11.4% / 21/30 / 4/20 remain on 1b5ab3e. UnifiedResult JSON for those runs is gitignored.
• Eval-protocol-v2 staged runner and worker slot keep-alive #186 fully contains docs: n=1540 product-recall handoff for next-cycle planning #185. Eval-protocol-v2 staged runner and worker slot keep-alive #186 and S5: versioned RRF retrieval for search and /recall #187 overlap only on worker keep-alive + current_database(); land Eval-protocol-v2 staged runner and worker slot keep-alive #186 first if merging.
• Do not merge leftover-covering PRs S1–S5 compiler/answer/alias/KU (split from mixed S0 PR) #133 / S0–S5 SOTA execution increments (stratified harness + compiler/answer/hop/LME) #131 / fix(recall): instrument-purpose leftover covering over help flood #143 / P29: demote single-token attended compiler echoes (ranking only) #145.
• Next eval spend, if approved: E0 (~$0) then E2 (new tenant, dual-lane 180), not another leftover-covering 180 and not LME-500.