nx_dup_source_check_v3.nx

ORPHAN FORK -- NOT THE FILE THAT RUNS. Adjudicated 2026-08-16 by measurement, not preference: * ZERO references to "nx_dup_source_check_v3" anywhere in buildroot -- nx_shelltool over 56,497 files, corpus_complete=1. Nothing imports it, no registry or conf names it. * The REGISTERED tool's banner ("scope DISCOVERED; allocation O(pairs) not O(files)") is emitted by nx_dup_source_check.nx, NOT by this file's header text. The name that ships is the SHORT one. ==> nx_dup_source_check.nx is CANONICAL. This file is dead weight kept only because retiring it needs nx_retire_path (rename-only, reversible), which the seat that found this could not call. ACTION: `nx_retire_path retire buildroot/runtime/_hdl_build/nx_dup_source_check_v3.nx`. A THIRD variant exists laptop-side (nxc2/runtime/_hdl_build/nx_dup_source_check_v3.nx) whose anchors match NEITHER NAS copy -- three forks of one organ. That is why its ARENA-OVERRUN went unfixed for so long: any seat repairing it had a one-in-three chance of editing a copy nothing executes. The slot-bound fix below was applied here anyway, so the forks cannot diverge FURTHER on this defect. ★A DUPLICATE-DETECTOR THAT IS ITSELF TRIPLICATED CANNOT REPORT ITS OWN CONDITION. nx_dup_source_check.nx v3 -- ANTI-CLOBBER cross-tree source-dup detector. 🚨v3 EXISTS BECAUSE v2 WAS AN OOM BOMB THAT TOOK THE NAS DOWN (2026-07-20). Root cause, stated plainly: v2 widened the scan from 2 hardcoded dirs to ALL discovered source trees (78 pairs) -- a good change -- but REUSED v1's `exists()` helper UNCHANGED, and that helper did `sys_mmap(4096)` PER CALL and never unmapped. v1 called it ~15k times (ONE pair) = ~60MB = survivable. v2 called it ~250k times = ~1GB LEAKED PER RUN, on a cron. Result: NAS userspace starved -- kernel still SYN-ACKs but sshd gives no banner, HTTPS returns status=0, even DSM stops answering. ★THE LESSON, so it is never repeated: CHANGING A CALLER'S FAN-OUT RE-COSTS EVERY CALLEE. A 2->78 pair change is a 17x amplifier on an O(files) helper. Reviewing only the code you WROTE is not enough -- you must re-cost the code you INHERITED AND AMPLIFIED. And note what the v2 selftest could NOT do: it was GREEN 4/4 because fixture dirs hold 2-3 files, so it tested the ALGORITHM at scale and never the RESOURCE at scale. A scale-law tooth that asserts only correctness-on-big-input is HALF a tooth; the other half is a RESOURCE bound. v3 FIX: every path buffer is allocated ONCE PER PAIR (or once per walk) and passed down, never per file. Allocation profile is now O(pairs + dirs), not O(files), and it is DECLARED in the output envelope so the property is inspectable rather than assumed. Detection semantics and the emitted " DUP basename: " line are UNCHANGED (rule 19) -- nx_favela_census counts that exact substring out of knowledge/status/dup_source.log. expect_exit:0 when clean.