nx_galx_cidput.nx

nx_galx_cidput.nx -- NATIVE incremental insert into the sovereign cid->path store. No TSV. Operator 2026-08-07: *"we shouldnt be using tsvs thats probably why it should be nishi native"*. Correct, and the estate had already said so: nx_galx_cid_index's own header reads *"The .idx + .blob ARE the runtime store; the TSV is only a one-time build input (and can then be retired)."* It was never retired -- because the store shipped READ-ONLY. gs_cidindex_build() can only build the whole thing FROM a TSV, and there was no way to add one entry. So every addition fell back to appending 70 bytes of text to a 34MB file and recompiling all 235,445 rows. ★★★★★★ A STORE YOU CAN ONLY BUILD, NEVER APPEND TO, KEEPS ITS OWN LEGACY INPUT FORMAT ALIVE FOREVER -- the "retired" format is load-bearing precisely because the replacement has no writer. ★ THE MISSING PRIMITIVE WAS NOT THE INDEX, IT WAS INSERT. The format already supports O(1) insert; nobody had written it. Open-addressed, fixed buckets: idx: [8 n_buckets][8 n_entries] then n_buckets * 24: [8 cid_hash][8 path_off][8 path_len] (big-endian) blob: concatenated path bytes So an insert is: append the path to blob, linear-probe to an empty bucket, write 24 bytes, bump n_entries. Three small writes at known offsets -- sys_lseek + sys_openat_rdwr, no file rewrite. ⚠LOAD FACTOR IS A HARD LIMIT, NOT A SUGGESTION. n_buckets is fixed at build time; an open- addressed table that passes ~0.5 degrades toward O(n) and at 1.0 never terminates its probe. This REFUSES past the builder's own nent*2 <= nb rule and names the remedy (rebuild, which re-sizes), because ★ A GUARD THAT CANNOT BE SATISFIED PRODUCES A BYPASS. license_tier: ORIGINAL