nx_role_store.nx

nx_role_store.nx -- LIB: the role-coordination layer in the SOVEREIGN seg-store (knowledge/store/roles-*), replacing the knowledge/roles/*.tsv files. Operator 2026-06-22: "migrate and get rid of these tsvs just nishi ecosystem." NO TSV. Same append-only-event-log + mutable-status-row shape as the .tsv queues, but the bytes live as content-addressed records in the crash-safe seg-store (nx_seg_store), NOT a flat .tsv file. CHANNELS (a channel = a logical .tsv replaced): pmcoord <- knowledge/roles/pm_coord.tsv (PM routing audit log, append-only) <role>q <- knowledge/roles/<role>_queue.tsv (per-role intake queue; row status is MUTABLE: PENDING->APPLIED) <role>l <- knowledge/roles/<role>_ledger.tsv (per-role apply ledger, append-only) RECORD SCHEMA inside a channel (mirrors nx_reader_census_store's keyed model): <ch>:n -> the channel row COUNT (decimal). rewritten each append (last-wins). <ch>:<seq> -> one row, 0-based seq. VALUE = the exact TAB-joined row bytes (the old .tsv line MINUS the trailing newline). A row's CURRENT status = the LATEST version of its <ch>:<seq> key (seg-store last-wins) -- so rq_pump flips PENDING->APPLIED by ss_add-ing a new version, no file rewrite. CONCURRENCY: every mutation holds an EXCLUSIVE flock on RS_LOCK across the read-count -> commit window (mirrors nx_ws_cas / pms_register's wlock) so concurrent submits/records never collide on the segid. HONEST BOUND (no silent cap, rule #11/#21): each commit adds ONE segment; ss_manifest + ss_scan cap at 256. rs_append REFUSES LOUD (returns -2) once the live segment count reaches RS_SEG_MAX, BEFORE staleness can occur. Compaction (ss_compact folds segments->1) is the next rung to lift the bound; it is NOT auto-driven here because the 1+count segid scheme would reuse archived segment ids -- a monotonic-id rung must land first. license_tier: ORIGINAL