fix(supply-chain): re-sync Cargo.lock to unblock the vet gate, and trust ewf's publisher

[h4x0r]

Two commits, both supply-chain hygiene. The second is unverifiable without the first, which is why they share a PR — but they are separate commits and read independently.

1. Cargo.lock re-sync — the vet gate has not been running at all

cargo vet shells out to cargo metadata --locked, and that has been failing outright:

ERROR   × `cargo metadata` exited with an error:
  │ error: cannot update the lock file ... because --locked was passed to prevent this

So the gate aborts before it evaluates a single crate. It has not been catching anything.

Root cause: PR #3 (ba519d9, scope OpenDAL backends to evidence-transfer targets). It cut 72 lines from Cargo.toml, dropping the cloud/object-store backends and their feature closure, and never regenerated Cargo.lock. The lock has carried the pre-#3 dependency graph ever since.

This is a re-resolution, not an update

No cargo update was run. Cargo keeps every pin the manifest can still reach and drops the ones it cannot. Measured on the (name, version) pair sets:

old	955 pairs
new	805 pairs
removed	150
added	0

The new lock is a strict subset of the old one — verified set-theoretically, not by reading the diff.

Nothing moved forward. The 16 packages that appear to have "changed" each lost a duplicate older major once its only consumer left the graph:

rand 0.7.3 · http 0.2.12 · hyper 0.14.32 · tower 0.4.13 · prost 0.12.6 and 0.13.5 · memmap2 0.5.10 · wasi 0.9.0 · rand_core 0.5.1 · sync_wrapper 0.1.2 · http-body 0.4.6 · tokio-rustls 0.24.1 · memoffset 0.7.1

One inversion worth naming explicitly: windows 0.62.2 was dropped and 0.52.0 retained — again because the 0.62 consumer left, not because anything downgraded.

Whole trees gone with their roots: axum, compio, cacache, bson, etcd-client, hickory-resolver, librocksdb-sys, bindgen.

Because nothing is added and nothing advances, this cannot introduce a new major, cannot raise the MSRV, and cannot bring in a licence cargo deny has not already seen.

The 7 +name = lines in the diff are relocations, not additions — packages re-emitted at a new offset after their neighbours were deleted.

2. ewf — trust the publisher instead of claiming we authored it

supply-chain/config.toml carried:

[policy.ewf]
audit-as-crates-io = false

That is ADR-0018 case 1, which means "ours and a workspace member". ewf is ours but it is not a member here — blazehash pulls it from crates.io like any other dependency. The policy told cargo-vet to skip auditing it on the strength of a claim that was not true.

The fleet's common defect is the opposite one: repos exempt their own crates and understate their posture. This entry overstated it, which is the direction that costs something — a compromised or substituted ewf release would have walked through the gate unexamined.

Verified against the registry before trusting. ewf has exactly one owner and every published version is theirs:

owners:  h4x0r | Albert Hui | user
0.4.7 -> published_by: h4x0r
0.4.6 -> published_by: h4x0r
...

So this is ADR-0018 case 2:

$ cargo vet trust ewf h4x0r --criteria safe-to-deploy

yielding a claim that is actually true — we trust this publisher, bounded to their user-id and a dated window:

[[trusted.ewf]]
criteria = "safe-to-deploy"
user-id = 347968 # Albert Hui (h4x0r)
start = "2026-03-05"
end = "2027-08-02"

It does not assert a human read the source. That is what a certify record would claim, and what audit-as-crates-io = false implied.

cargo-vet removed the now-redundant [[exemptions.ewf]] 0.2.3 itself — case 2 outranks case 4, and keeping both would leave the weaker mechanism carrying a load it no longer bears.

The imports.lock churn is fallout from commit 1: cexpr's publisher record drops out because bindgen left the graph, and one upstream mozilla entry lost an inline comment on re-fetch.

Verification

No honest RED commit exists for either change — the failure is a tooling abort, not a behaviour a test can assert. Before/after output stands in its place:

BEFORE  ERROR × `cargo metadata` exited with an error:
        error: cannot update the lock file ... because --locked was passed

AFTER   Vetting Succeeded (170 fully audited, 5 partially audited, 628 exempted)

gate	result
cargo build	pass
cargo test	pass — 214 passed, 0 failed, 10 ignored
cargo clippy --all-targets -- -D warnings	pass
cargo fmt --check	pass
cargo vet	pass (was aborting)
cargo deny check	FAIL — pre-existing, see below

Known red this PR does not fix

cargo deny check fails, and fails identically on pristine origin/main — reproduced on a detached worktree at 366a799:

error[vulnerability]: Stores can mix up type indices between engines
    wasmtime 25.0.3 — RUSTSEC-2026-0222
    Solution: >=24.0.12,<25.0.0 OR >=36.0.13,<37.0.0 OR >=46.0.2,<47.0.0 OR >=47.0.3
    wasmtime v25.0.3 <- yara-x v0.9.0 <- blazehash v0.2.6
warning[yanked]: spin 0.9.8
advisories FAILED, bans ok, licenses ok, sources ok

The lock re-sync cannot have introduced it — the new graph is a strict subset of the old. Closing it means moving off yara-x 0.9.0, which pins wasmtime 25.x. That is a dependency decision, not a drive-by, so it is flagged rather than attempted here.

Follow-ups deliberately not bundled

• 157 orphaned vet exemptions. The prune leaves cargo vet warning about unnecessary exemptions: 786 entries before, 629 after cargo vet prune. It is a WARN, cargo vet --locked still exits 0, nothing blocks. Left out to keep this diff readable.
• ewf is caret-trapped. Cargo.toml:112 pins ewf = { version = "0.2", optional = true }; the lock sits at 0.2.3 while the registry is at 0.4.7 (2026-07-25) — a layer-1 freshness trap cargo update cannot cross. Widening it adds packages and would destroy the "strict subset, nothing added" property that makes commit 1 safe to review at a glance. It wants its own PR, plus a check that 0.2 -> 0.4 is API-compatible.
• ImageFormat::detect dispatches on file extension (src/forensic_image/mod.rs:19-37) where the fleet standard is content-sniffing. That is an ADR-0011 architectural call, reported not attempted.

🤖 Generated with Claude Code