nx_memcliff.nx

nx_memcliff.nx -- SOVEREIGN owner of the MEMORY.md byte budget. Replaces memindex_enforce.py (Stop hook) and the eviction half of memory_cliff_check.ps1 (SessionStart/PostToolUse) with ONE nishi organ, so no interpreter sits in the hook path. WHY AN ORGAN AND NOT A SCRIPT ----------------------------- MEMORY.md is injected into every session; past a byte limit the harness DROPS THE TAIL SILENTLY -- no error, no marker. The dropped lines are exactly the ones that would have told you they were dropped. The rail-side enforcer cannot own this file (it runs NAS-side; MEMORY.md is laptop-local, nx_memindex_put -> `no-md-file`), so ownership has to live here, next to the file. WHAT WAS ACTUALLY BROKEN -- MEASURED 2026-08-06, not assumed ------------------------------------------------------------ The Python enforcer was NOT missing: it had run 64 times (64 eviction stamps in the overflow) and had compacted the file at 13:47 the same day. It had JAMMED: "22013 bytes over budget but NO evictable entries found -- refusing to guess" Root cause, measured: its rule NEVER EVICT THE NEWEST DAY pinned 8 of 8 candidates, because every evictable entry carried today's date. Survivors after the pin: ZERO. ★LAW: A PIN THAT CAN COVER 100% OF ITS CANDIDATES IS AN OFF SWITCH. The rule that protects today's record disables enforcement precisely on the heaviest days -- the file grew 4.2KB in the two hours after that successful compaction, with nothing left able to evict it. The fix is not to drop the pin (it earned its place: a pure oldest-first cut once ate the current session's own record) but to make it YIELD: keep the newest `keep_newest_day` entries of the newest day and release the rest, rather than pinning the whole day unconditionally. SECOND MEASURED DEFECT -- the enforcer could not measure its subject in the unit the limit uses. MEMORY.md is CRLF (63 CRLF, 0 bare LF). Python read it in text mode, which collapses CRLF to LF, so it measured 22,013 against a file of 22,076 bytes -- UNDERSTATING by exactly the line count. The harness cliff is a property of the bytes on disk, so every decision was made against a number that was never the file's size. ★LAW: A BUDGET ENFORCER THAT CANNOT MEASURE ITS SUBJECT IN THE UNIT THE LIMIT USES IS GUESSING. ⚠WHAT I FIRST WROTE HERE AND THEN DISPROVED: I claimed the two owners also flipped the file's line endings against each other (py writing LF, the ps1 writing CRLF). THE EXPERIMENT REFUTED IT -- a CRLF fixture through the oracle came out CRLF=26 / bare-LF=0, because Python's text-mode WRITE translates back on Windows, exactly undoing the read. The round trip is lossless; only the measurement was wrong. Recorded rather than quietly deleted, because a plausible mechanism I had not run is precisely the kind of premise that spends like evidence. This organ works in RAW BYTES and reconstructs each line with its original terminator, so it is byte-exact by construction rather than by a translation that happens to cancel out (verified: the same CRLF fixture out at CRLF=26 / bare-LF=0).