fix(security): delete the plaintext LFS mount rather than gate it

[forkwright]

The finding inverted its own description

MountPlan::Plain mounted LFS at the userdata partition head with no EncryptedBlockDevice, and formatted it on InvalidSuperblock. Its comment called it a "legacy dev/transition path", and #866 asked for it to be removed from production or build-gated as non-production.

It was reachable only in production. mount_plan cannot reach that arm without secure_boot_ok, and kinit sets that at exactly one place — gated on BOOT_KEY_IS_PRODUCTION:

if crate::secure_boot::BOOT_KEY_IS_PRODUCTION {
    state.secure_boot_ok = true;

A dev-anchor build boots degraded with it false, deliberately, so the committed public dev key cannot establish trust on a device. Host tests always build with the dev anchor. QEMU never verifies at all.

So no dev build, no host test and no QEMU witness could ever take that arm, and the only configuration that could was the one where a plaintext persistent root is forbidden. A compatibility path that exists in no configuration but the one it must not run in has nothing left to be compatible with — which is why this deletes it rather than taking the gating option the issue offered. Gating would have produced a branch no build can execute and no test can reach.

What changes on a device

An unprovisioned device whose first-boot setup did not produce a verified passphrase — skipped, interrupted, or unable to reach the operator — now falls back to the initramfs instead of creating a plaintext root. It has never held a secret, so it has no userdata to lose by declining to mount.

The property is structural now, not asserted

MountPlan loses its third variant, so there is no plaintext answer to return. The new exhaustive test walks the whole input space — every combination of emmc_ok, secure_boot_ok, all five PreambleLoad states and passphrase_ok — rather than the states someone remembered to enumerate, and requires every Encrypted result to carry a derived key on a verified boot over a readable preamble.

One thing clippy found that was worth more than silencing

With the arm gone, preamble became an unused parameter — the decision no longer consulted it. _preamble would have compiled.

Instead the check is now stated: Encrypted requires passphrase_ok and a preamble of Provisioned or Unprovisioned. Today that is implied — boot_passphrase_plan answers Skip for the other three, so no key can exist for them — but the implication lives in a different function. The cost of it ever ceasing to hold is mounting the payload of a device whose preamble could not be read, which is exactly what #621 established must never happen. Three new assertions pin that unknown stays locked even if a key somehow exists.

Also retires the plain arm's half of #360

The encrypted mount distinguishes a never-formatted payload from a damaged one via provisioned_this_boot. The plain arm had no such marker and formatted anyway; the comment explaining why that was tolerable is gone with the code it excused.

Verified

Nine-configuration scripts/kernel-clippy.sh clean, cargo fmt --check clean. Full kernel suite on the host target: 2463 tests, 43 failures — exactly the unmodified-main baseline, compared name-by-name against a captured baseline set rather than by module prefix. 129 lines removed from kinit.rs; docs/capability-inventory.toml updated so it no longer describes a plaintext path that does not exist.

Closes #866