mirror of
https://github.com/hanzoai/cloud.git
synced 2026-11-17 16:36:24 +00:00
Encrypting() answers whether a master key is present. It was being read as whether this build can encrypt, and the comment at the boot posture asserted that every backend can. Measured, an ordinary `go build` does not: cgo links plain SQLite unless the build says -tags libsqlite3 against libsqlcipher. On such a build the key is present, so the boot announced "data-plane encryption ACTIVE", every store was written with the plaintext SQLite header, and the SECOND boot died in migration needing a SQLCipher function it did not have. Between those two moments it had written data to disk in the clear while saying it had not, which is the worst shape a privacy claim can take. The pure-Go backend has an honest error for exactly this; the cgo-without-the-library path had none. Capable answers the second question by DOING it: mint a keyed store, write to it, and ask the same classifier the open path asks. A probe that read a build tag or looked for a symbol would be a second opinion about the one thing that matters. It is enforced at Open, not at boot, because boot is not early enough — the posture check runs after the gateway's own store is open, and that one file was already on disk in plaintext when I put the check there. At the open seam no ordering anywhere can get past it: measured, zero stores reach the disk before it refuses, against one before. Six test files named a key to get a store. That is a request to encrypt, and on a build without the library they were writing plaintext while looking like they exercised an encrypted one. They ask for the development path now, which is what they actually wanted, and they pass on either build. The bench/self lane says this out loud rather than burying it: running from source is free and takes a second, and an encrypted data plane costs you a build with the library linked. Co-authored-by: Hanzo Dev <[email protected]>