nx_fp_backfill.nx

nx_fp_backfill.nx -- BACKFILL THE fp: FINGERPRINT ROW OVER ALREADY-INGESTED DOCS. WHY: nx_web_crawl_step now persists fp:<cid> (the simhash it always computed and used to discard), so intra-host near-duplication -- the real doorway signature -- is a property of the index. But that is FORWARD-ONLY: every doc ingested before 2026-08-16 has no fp: row, so a duplication census would silently measure only the newest slice and report it as the corpus. This closes that gap. NO NETWORK. The text is already stored; this reads doc:<cid>, hashes it, writes fp:<cid>. A source that died years ago costs nothing here -- which is the whole argument for an entity corpus over a page index: capture once, and every later question is local. MEMORY SHAPE, STATED because the sibling's is different: nx_web_shard_compact was OOM-killed twice (bytes-in=1883885124, 2026-07-25) because it MATERIALISES a merged segment in RAM. This does not -- it holds one doc at a time and emits a ~20-byte row, so its bound is on RUNTIME, not memory. BOUNDED AND RESUMABLE, never a silent runaway: argv[1] max REQUIRED and must be > 0 -- refuse rather than default to unbounded argv[2] skip resume offset into the walk order (default 0) argv[3] commit literal "commit" to write; DEFAULT IS A DRY RUN that changes nothing PROGRESS IS PROVABLE: every run prints scanned/had_fp/written/capped. The sibling's range-fold note is the warning -- a bounded sweep with the wrong selection rule runs forever WITHOUT ADVANCING, so "did this converge" must be a number, not a feeling. capped=1 says the run stopped at max, i.e. the corpus is NOT yet fully covered and the next run needs skip=<scanned+skip>. license_tier: ORIGINAL