nx_planepeek.nx

nx_planepeek.nx -- READ ONE RAW KEY OUT OF A SEG-STORE PLANE. Read-only, writes nothing. WHY IT EXISTS. The debt- plane is refusing every add/eat because ONE row is reachable past the declared q:n (nx_plane_check: rows=3696 identical=3696 declared_qn=3696 beyond_qn=1), and nx_plane_repair correctly refuses to reconcile it: "an older seeding generation -- reseeding would revive eaten rows and duplicate live ones. That needs a generation diff, not a reconcile." To DO that generation diff somebody has to look at the row. And nothing in the estate could: nx_store_put reads its own N-column planes, nx_plane_check verifies round-trips, nx_store_hist is a column histogram -- none of them will hand you the value at an arbitrary key. ★★★★★A PLANE YOU CAN VERIFY AND REPAIR BUT CANNOT READ A SINGLE KEY OUT OF FORCES EVERY DIAGNOSIS TO BE A GUESS. I made two wrong guesses about this exact row before building this. It is deliberately the smallest thing that closes that gap: resolve a key through the SAME production ss_get every reader uses, print the bytes, exit. No writes, no repair, no opinion. =========================================================================================== THE FALSE-ABSENCE DEFECT (debt 1787282525, sev 6, fixed 2026-08-21). =========================================================================================== v1 printed exactly one word for every non-hit: `if ss_get(...) != 1 { MISS }`. ss_get's OWN contract (nx_seg_store.nx: "1=found, 0=tombstoned, -1=absent") already carried a state v1 threw away, and the scan underneath it carried three more: 1. ss_get==0 -> the key EXISTS and was explicitly DELETED. v1 said MISS. A tombstone is a POSITIVE fact about a key, and reporting it as absence inverts its meaning. 2. no manifest -> ss_manifest_dyn returns 0, ss_scan returns 0, ss_get returns -1. So a MISTYPED PREFIX printed the same MISS as a real absence. Nothing could tell them apart. 3. a segment NAMED in manifest.txt but UNREADABLE on disk -> ss_readall sets sz=-1 and ss_scan_seglist's `while i + 9 <= sz` loop never executes. The segment contributes ZERO rows and NO error. A plane half of which could not be opened answered a confident "absent". 4. a caller passing a COLUMN VALUE where the plane's keys are q:<n> (every N-col plane written by nx_store_put) got MISS, which reads as "your row is gone" rather than "wrong grammar". FIVE distinguishable conditions, ONE word, and that word is a NEGATIVE ASSERTION. Measured cost: a seat nearly filed "the enqueue lane is dead" on it, and a build-refusal message shipped `verify with nx_planepeek deployq- BQ-<target>` -- pointing every caller at a false-absence generator. ★★★★★★AN UNBOUNDED SEARCH THAT GIVES UP MUST NOT REPORT THE SAME WORD AS AN EXHAUSTIVE ONE THAT FOUND NOTHING -- otherwise every quote of it silently rewrites the evidential status of the claim that quotes it. ⇒ MISS now means ONE thing: the plane resolved, EVERY declared segment was read, the key is not there. Everything else gets its own word and its own exit code, and every answer -- hit or miss -- prints the coverage it was computed from. ★PRESENCE NEEDS ONE WITNESS; ABSENCE NEEDS PROVEN COVERAGE. So the HIT path stays a cheap point