nx_memindex_put.nx

nx_memindex_put.nx -- the ATOMIC, SLUG-KEYED upsert for the HAND-CURATED half of the memory index (F826 / debt seq273; the contention half of the class whose budget half shipped as F825). WHY THIS EXISTS. The COINDEX block is a conflict-free rail: sessions APPEND to a journal and one writer splices the derived block. The curated pointer block below it had no such path, so every session mutated the shared index with a read-modify-write Edit -- which raced siblings FIVE times in one session and once produced a MANGLED CONCATENATED LINE. The defect is structural: a stale read plus a blind overwrite. This organ removes both -- it re-reads the file INSIDE the lock and rewrites exactly ONE slug-owned line, so a sibling's line can never be clobbered by a stale view. WHY NOT MAKE THE CURATED BLOCK JOURNAL-DERIVED TOO? Because it is genuinely EDITORIAL: the pointers are grouped thematically (DEPLOY/MGMT, COORDINATION/PM, ...) by human judgment, and deriving it per-slug would destroy that grouping. So authorship stays human; only the MUTATION becomes atomic. nx_memindex_put <journal> <md> <slug> <line> [ovfl] Locks <journal>.lock -- the SAME lock nx_coindex append and nx_memindex_emit take -- so put, append and splice are fully serialized against each other, not merely against other puts. ---- BUDGET + AUTO-EVICT (added 2026-07-20 for debt seq255 sev7 / F847; ADDITIVE, rule-19) -------- The upsert above kills the CLOBBER half of seq255. It does NOT kill the SCALE half, which the debt row names precisely: "unbounded growth (20.4->21.2KB mid-session, over the 17.1KB soft target, siblings add faster than anyone can hand-compact) + no session can safely compact". A WARNING does not fix that -- warnings are addressed to a librarian who, by the row's own account, cannot keep up. So with an [ovfl] file the zone becomes a true RAIL, symmetric with the COINDEX block: the budget is DERIVED (whole-file target minus everything outside the zone, so it self-adjusts as the rest of the file moves) and the OLDEST lines auto-evict into the overflow file's CURATED-OVFL block. Orientation matters and is not arbitrary: this organ APPENDS new lines at the END of the block, so the zone is oldest-first and ci_split_at_budget -- which keeps the newest SUFFIX -- is exactly the right splitter, reused rather than re-derived. NO-LOSS beats budget: if the overflow file is missing or markerless, NOTHING is evicted and the organ says so. Omitting [ovfl] preserves the old behavior byte-for-byte (warn, never evict). Structural lines (thematic groups, LAWS, DOCTRINE, the gotcha tail) live OUTSIDE the markers and are unreachable here, so the round-4 injury -- silently dropping the file's own laws -- cannot recur. Each managed line carries an invisible owner marker `<!--s:SLUG-->` at its end. Upsert = replace the line carrying that marker IN PLACE (position preserved, so the editorial grouping survives), else append just before the CURATED:END marker. FAIL-CLOSED: missing/misordered CURATED markers -> file untouched and says so (never guesses where to write). ATOMIC: tmp + fsync + renameat, so a concurrent reader sees whole-old or whole-new. IDEMPOTENT: putting identical text twice leaves the file byte-identical. It also REPORTS the resulting curated size against the whole-file target, so a writer learns it is