Repository navigation
feat(mfile): T26 — Project the spec and digest it the same way in client and daemon - #3373
Conversation
Add the Gatehouse §6.3.2 Box Spec projection and its digest: SHA-256 over the record's RFC 8785 JCS bytes, encoded base64url (BOX-071). Expansion now writes each store reference's upstream and inject from the client's [secret-store-rules], and fills the omitted github token mode, network.quic443 and network.bep.steering before the projection. Expansion takes no host enrollment input, so enrolled and un-enrolled hosts expand to the same bytes and digest (BOX-142).
The create request carries the client's projection digest in its attrs. The daemon recomputes the digest from the spec it received and stores it beside the client's in the record (BOX-071). The create span and the daemon log name both digests. A daemon digest that a client sends is dropped.
Writing the filled quic443, steering and token mode into the expanded spec changed the `min box spec` output that a minimal crate test pins. The projection now fills them for the client and the daemon alike, so the digest still binds the effective values and the spec bytes keep only what a layer set.
…e-spec-and-digest-it-the-same-way-in-c
|
Important Review skippedBot user detected. To trigger a single review, invoke the
|
|
Reviewed: no findings; approval withheld, 12 threads unresolved at APPROVE withheld: 12 unresolved threads — minimal-arch[bot] crates/mfile/src/projection.rs, minimal-arch[bot] crates/mfile/src/expand.rs, minimal-arch[bot] crates/mfile/src/expand.rs, minimal-arch[bot] crates/mfile/src/projection.rs:377, minimal-arch[bot] crates/mfile/src/projection.rs, …. Resolving the last thread releases the approval on this head. Resolve the open threads. Resolving the last one releases the approval on this head; no new review is required. Verdict APPROVE, posted as a comment: the App cannot review its own pull request; the check and the label carry it. Clobber: none found. |
|
Round 1 of 10 on the runtime's review findings: the coder changed nothing. Revise round 1: I made no code change, so this round has no commit. The one thread is invalid for this task. Its first claim describes behaviour that the spec requires. Its second claim is true, but the fix is outside the task's layers. Production clients do not fill Layers.store_rules from the client's [secret-store-rules], and they do not call set_client_projection_digest. That wiring is in crates/minimal/src/cmd/box_spec.rs and crates/minimal/src/cmd/boxes/create.rs. It needs a follow-up task with those files in its layers. Until it lands, the daemon stores only its own digest for boxes that min creates. I did not run the test: lines this round, because nothing changed. Review threads the coder judged:
|
There was a problem hiding this comment.
arch-review: CONDITIONS · 6 conditions · the full ruling
The daemon side conforms. It recomputes the digest, stores both digests in the record, and names them in the span and the log. The projection record does not match Gatehouse §6.3.2. secret_grants copies each entry's own keys and adds name (C1), and upstream is a {hosts} table where the design has a list (C2). ingress_ports holds objects (C4), and nesting copies the whole table (C5). Expansion drops a project-written upstream/inject, but box.toml lines 200–202 make it an exit-3 error (C3). The client never computes or sends its digest, so file that work as a residual with an issue (C6).
Ruled at beeeae4 against gominimal/arch@15b603e6afb8 by claude-opus-5-5.
| Json::Object(map) => map, | ||
| other => BTreeMap::from([("value".to_owned(), other)]), | ||
| }; | ||
| grant.insert("name".to_owned 10000 (), Json::Str(name.clone())); |
There was a problem hiding this comment.
C1 · Section 6.3.2 lines 664-668 name the grant field env, not name. They also give a fixed key set for each kind of grant. Build each grant from those keys only. Do not copy the entry's own keys.
There was a problem hiding this comment.
Addressed in 7eec2ca508d8: secret_grants() now builds each grant from its kind's §6.3.2 key set with the entry name as env: store refs carry env/source/store/id/upstream/inject, brokered grants env/scopes/mode, no source.
| grant.insert( | ||
| "upstream".to_owned(), | ||
| toml::Value::Table(toml::Table::from_iter([( | ||
| "hosts".to_owned(), |
There was a problem hiding this comment.
C2 · The design writes upstream as a list of authorities (box.toml line 208, section 6.3.2 line 667). Write upstream = [..] here, not { hosts = [..] }. Change the store_secret() fixture in box_manager/tests.rs to the same shape.
There was a problem hiding this comment.
Addressed in 7eec2ca508d8: write_store_authorities() writes upstream = [..] as a list of authorities. The store_secret() fixture in box_manager/tests.rs uses the same shape, and the test asserts it.
| if grant.get("source").and_then(toml::Value::as_str) != Some("store") { | ||
| continue; | ||
| } | ||
| grant.remove("upstream"); |
There was a problem hiding this comment.
C3 · box.toml lines 200-202 say that an entry that sets upstream or inject is a validation error (exit 3). Refuse it in expand() before this function runs, and remove this silent drop.
There was a problem hiding this comment.
Addressed in 7eec2ca508d8: expand() now refuses a store reference that a layer gives upstream or inject (InvalidSpecKey, exit 3, names secrets..) before write_store_authorities(). The silent drop is removed.
| /// spec's order. It reads the `box.toml` list of `{ proto, port }` and | ||
| /// the `port_mappings` of a v0 spec, whose port in the box is | ||
| /// `internal_port`. | ||
| fn ingress_ports(spec: &toml::Table) -> Json { |
There was a problem hiding this comment.
C4 · Section 6.3.2 line 657 gives "ingress_ports": [8080]. Project each port as its number, not as a { port, proto } object.
There was a problem hiding this comment.
Addressed in 7eec2ca508d8: ingress_ports() projects each admitted port as its number. projection_is_the_6_3_2_field_set asserts [8080].
| )]); | ||
|
|
||
| let commit = at(spec, &["upstream", "commit"]).and_then(toml::Value::as_str); | ||
| let nesting = match spec.get("nesting") { |
There was a problem hiding this comment.
C5 · Section 6.3.2 line 663 gives nesting as five bounds: allowed, max_depth, types, providers and dynamic_providers. Project only those keys. Do not copy the whole [nesting] table.
There was a problem hiding this comment.
Addressed in 7eec2ca508d8: project() now projects nesting as exactly allowed, max_depth, types, providers and dynamic_providers. A test case shows that an extra [nesting] key leaves the digest unchanged.
| /// `box_type` (BOX-071). Its [`digest`](crate::projection::Projection::digest) | ||
| /// is the one the client sends with the spec. | ||
| #[must_use] | ||
| pub fn projection(&self, box_type: &ResolvedType) -> crate::projection::Projection { |
There was a problem hiding this comment.
C6 · No client code calls this function or set_client_projection_digest, and the client Layers never sets store_rules. Section 6.3.2 line 649 says the client computes the projection. File that client work as an issue, and name it as a residual in the PR body.
There was a problem hiding this comment.
Addressed in 7eec2ca508d8: The client half is outside this task's layers and is filed as #3377. It is to be named in the PR body as a residual of T26. No code change in this round.
|
/revise Arch conditions (arch_rulings 1230, ruled at beeeae4), in full:
C6 is filed as #3377. Name it in the PR body as a residual of T26. |
…e-spec-and-digest-it-the-same-way-in-c
Build each secret grant from its kind's key set, with the entry name as env. Project ingress ports as numbers and nesting as its five bounds. Write a store reference's upstream as a list of authorities, and refuse a layer that sets a store reference's upstream or inject (exit 3).
The design writes a store reference's upstream as a list of authorities, so the fixture uses the same shape.
|
Round 2 of 10 on a /revise note, pushed as
Verify Converted to a draft: something above needs a person before review. Review threads the coder judged:
A created box's record holds both projection digests. The digest now covers the §6.3.2 record: each secret grant uses its kind's key set with the entry name as env, ingress ports are port numbers, nesting is its five bounds, and a store reference's upstream is a list of authorities. A minimal.toml that sets a store reference's upstream or inject is refused with exit 3. Notes from the agent: Revise round 2, conditions C1 to C5 done. C6 is a PR-body item, and I cannot edit the PR body. Please name #3377 there as a residual of T26: the client half of BOX-071 (crates/minimal/src/cmd/boxes/create.rs sends no expanded spec, sets no Layers::store_rules and no client projection digest). C1: the code has no |
There was a problem hiding this comment.
arch-review: PASS_WITH_CONDITIONS · 1 condition · the full ruling
The compare closes the earlier deviations. ingress_ports, nesting and secret_grants now have the shape that Gatehouse §6.3.2 gives. One gap remains: min never sends the client digest. Gatehouse §6.3.2 line 649 says the client and the node each compute the projection. No code in crates/minimal calls set_client_projection_digest, so a real create stores the daemon's digest alone. That work is outside this task's layers. Name it as a residual with its issue in the PR body, as C1 says.
Ruled at 7eec2ca against gominimal/arch@15b603e6afb8 by claude-opus-5-5.
|
|
||
| /// Sets the projection digest the client computed for the spec it | ||
| /// sends (BOX-071). | ||
| pub fn set_client_projection_digest(&mut self, digest: impl Into<String>) { |
There was a problem hiding this comment.
C1 · No code in crates/minimal calls this setter. put_spec in crates/minimal/src/cmd/boxes/create.rs sends the spec with no digest. Gatehouse §6.3.2 line 649 says the client computes the projection too. Add a residual to the PR body with its issue for the client wiring.
There was a problem hiding this comment.
Disputed — not addressed: The finding is true, but the fix is outside this task's layers. The client call to set_client_projection_digest belongs in crates/minimal/src/cmd/boxes/create.rs (put_spec), and this task does not name that file. Until then the record holds the daemon digest alone. The client wiring needs its own follow-up task and a residual in the PR body.
A client value that is not 43 base64url characters is dropped from the record, the span and the log, with a warning. The mfile doc on SpecDigestError now says a null reads as an absent key.
…e-spec-and-digest-it-the-same-way-in-c
|
Round 3 of 10 on the runtime's review findings, pushed as
Verify Converted to a draft: something above needs a person before review. Review threads the coder judged:
A created box's record holds both projection digests: the daemon's, recomputed from the spec, and the client's when the client sent one in digest form. Notes from the agent: Revise round 3. The daemon now stores a client projection digest only when it is 43 base64url characters. Any other value is removed from the record attrs, left out of the span and the log, and logged as a warning with its length only. The side-by-side test now uses a well-formed differing digest and checks four malformed values. The SpecDigestError::NotATable doc now matches the code: a null reads as an absent key. The residual from C6 (no client code computes or sends the digest yet) still stands. just clippy-strict was not run. |
|
Round 4 of 10 on a failed check suite: the coder changed nothing. Revise round 4: no code change. All five test: lines pass at the merged head 4a44bad. The failed test-kvm check: the excerpt shows only the job's script text and post-job cleanup, and it names no failed test. This box has no /dev/kvm, so I did not run the VM harnesses. No harness under crates/minvmd/tests reads record attrs, the projection digests or the new log lines. The daemon change only adds attrs and a log line, and it never refuses a create. So I found no link from this diff to the failure. A person must read the full test-kvm log, or re-run the job, to see if it is a flake. ci-linux-kvm-success failed only because test-kvm failed. Thread C1 (client wiring): the fix is in crates/minimal/src/cmd/boxes/create.rs, which is outside this task's layers. It needs a follow-up task. Add that residual to the PR body; the runner writes the body, not me. Review threads the coder judged:
|
|
Revise round 4 ran against |
|
/revise Merge |
|
Round 5 of 10 on a /revise note: the coder changed nothing. every commit this round put on the branch above 4a44bad is a merge of the base branch, so it carries no work of its own |
There was a problem hiding this comment.
minimal-review — approval withheld
Verdict COMMENT — APPROVE withheld: 9 unresolved threads — minimal-arch[bot] crates/mfile/src/projection.rs, minimal-arch[bot] crates/mfile/src/expand.rs, minimal-arch[bot] crates/mfile/src/expand.rs, minimal-arch[bot] crates/mfile/src/projection.rs:377, minimal-arch[bot] crates/mfile/src/projection.rs, …. No findings.
Panel: 3 of 3 personas ran (correctness, security, supply-chain).
There was a problem hiding this comment.
arch-review: PASS_WITH_CONDITIONS · 1 condition · the full ruling
The digest form check in 3267edf4 conforms. It keeps an untrusted client string out of the record and the log. Now a malformed client digest reads as none on the span and in ProjectionDigests, the same as a client that sent no digest. Gatehouse §6.3 check 5 and §6.3.2 line 672 treat a client and node that disagree as a refusal. BOX-072 must therefore know that the client sent a value. C1 records that state as malformed. C1 can stay open until BOX-072 lands. The client-side create call stays on residual #3377.
Ruled at 4a44bad against gominimal/arch@15b603e6afb8 by claude-opus-5-5.
| spec: Option<&serde_json_lenient::Value>, | ||
| verb_type: &str, | ||
| ) -> ProjectionDigests { | ||
| let client = config |
There was a problem hiding this comment.
C1 · This filter turns a malformed client value into None, the same state as a client that sent no digest. Gatehouse §6.3 line 624 refuses a digest that differs from the daemon's. Keep a distinct Malformed state in ProjectionDigests. Record client_projection_digest = "malformed" on the span and in the log, not "none". Then BOX-072 can refuse it.
There was a problem hiding this comment.
Addressed in 7ffedf837e00: projection_digests() now maps a malformed client value to ClientDigest::Malformed. The span and the info log record client_projection_digest = "malformed", not "none".
| /// received. | ||
| #[derive(Debug, Clone, Default, PartialEq, Eq)] | ||
| struct ProjectionDigests { | ||
| /// The client's digest, `None` when the client sent none. |
There was a problem hiding this comment.
C1 · This doc says None means that the client sent no digest. After 3267edf4, None also means that the client sent a malformed value. Change the type or the doc so that the two states stay distinct (C1).
There was a problem hiding this comment.
Addressed in 7ffedf837e00: The client fi
10BC8
eld is now a ClientDigest enum with Absent, Malformed and Digest states. Its doc says that the client sends a digest, or it gives the reason that there is none.
|
/revise Arch condition (arch_rulings 1260, ruled at 4a44bad), in full:
|
…e-spec-and-digest-it-the-same-way-in-c
A client projection digest that is not in digest form is now its own state, so BOX-072 can refuse it as a differing digest. The raw value stays out of the record, the span and the log.
|
Round 6 of 10 on a /revise note, pushed as
Verify Converted to a draft: something above needs a person before review. Review threads the coder judged:
A created box's record holds the client's projection digest beside the daemon's. A client value that is not in digest form is kept out of the record, and the create span and the daemon log name it as "malformed", not "none". Notes from the agent: Round 6 addresses arch condition C1. ProjectionDigests.client is now a ClientDigest enum with the values Absent, Malformed and Digest(String). record() writes "malformed" for a malformed value, and the raw value stays out of the record, the span and the log. The test now checks the "malformed" log line and checks that the raw value is not in the log. cargo clippy -p minimald --all-targets is clean. The client wiring is still a residual from earlier rounds: no code in crates/minimal sends the digest yet. |
There was a problem hiding this comment.
arch-review: PASS_WITH_CONDITIONS · 1 condition · the full ruling
The projection matches the Gatehouse §6.3.2 field set and digest, and the new ClientDigest::Malformed state correctly fails closed. One deviation stays open from the earlier heads. Section 6.3.2 line 672 says expansion fills the grant mode, quic443 and steering. This change fills them only inside projection::project, so the expanded spec bytes do not carry them. C1 moves the fill into expand and updates the pinned min box spec test. You can merge with C1 open when its issue is filed and named in the PR body.
Ruled at 7ffedf8 against gominimal/arch@15b603e6afb8 by claude-opus-5-5.
| // BOX-071: the store references carry the client's authorities, so | ||
| // the daemon projects the same `secret_grants` from the spec it | ||
| // receives. The omitted values the digest binds are filled by the | ||
| // projection itself (`projection::fill_effective`), in the client and |
There was a problem hiding this comment.
C1 · Section 6.3.2 line 672 says expansion fills the grant mode, quic443 and steering before canonicalization. Here the fill happens only in projection::project, so the spec bytes do not carry it. Call projection::fill_effective(&mut expansion.spec, root) after write_store_authorities, and replace this comment.
There was a problem hiding this comment.
Disputed — not addressed: BOX-071 says to fill before the projection is canonicalised so that the digest binds the effective values. projection::project runs fill_effective in both the client and the daemon, and fill_effective is idempotent, so both digests bind auto/dns/the token mode. Writing the fill into the spec bytes changes the min box spec output that a crates/minimal test pins. That file is outside this task's layers, so commit cfbbf30 moved the fill into the projection.
The daemon's admission check reads only box.type. A flat type string made the digest name a type that admission never checked.
|
Round 7 of 10 on the runtime's review findings, pushed as
Verify Converted to a draft: something above needs a person before review. Review threads the coder judged:
A created box's record holds both projection digests: the client's and the daemon's. The daemon now computes its digest for the same type that admission checks. Notes from the agent: Round 7. The mfile test lines ran in one nextest invocation together with type_facts_of_spec_reads_box_type_and_root, and all passed. I did not run just ci or clippy-strict. The expand.rs fill thread is judged invalid. The fill stays in projection::fill_effective, which the client and the daemon both run before canonicalisation. A filled spec changes the |
|
Round 8 of 10 on a failed check suite: the coder changed nothing. Round 8 changed nothing, so noop is true. Failed check 'tests': the job excerpt does not name the failing test. I ran the full mfile suite (169 passed) and the full minimal and minimald-rpc suites (1211 passed). I ran the full minimald suite: 1491 passed. One test failed (env::tests::task_entrypoint_carries_env_cwd_and_args) and two hung (env::tests::minenv_run_forwards_stdin_to_the_task, env::tests::task_run_pipes_stdin_through_the_in_box_helper). All three are in crates/minimald/src/env.rs, which this branch does not change. They pass when MINIMAL_OTEL is unset. This box sets MINIMAL_OTEL=off, and the test does not expect the otel prefix in the request. So the cause is the environment of this box, not the branch. I did not run a workspace-wide test, because no test: line names one. The CI failure can be a test outside these crates or a flaky test. A person must read the full job log to name it. Doctests: rustdoc is not on this box, so cargo test --doc did not run. Failed …[truncated] |
|
Revise round 8 ran against |
There was a problem hiding this comment.
arch-review: PASS_WITH_CONDITIONS · 1 condition · the full ruling
The compare conforms. TypeFacts::of_spec now reads the type from box.type only. The Box Spec carries the type there (box.toml lines 12–27), and admission reads only that key. Thus the digest names the type that admission checked (§6.3.2 line 672). The rest of the projection holds as it did on 7ffedf83. The one open item is C1: the min create path must send the client digest. #3377 tracks C1, and the merge can go ahead with C1 open. The agent-verify leg timed out after a test failure. That leg is open, and a person must check it before merge.
Ruled at 9e053ff against gominimal/arch@15b603e6afb8 by claude-opus-5-5.
|
|
||
| /// Sets the projection digest the client computed for the spec it | ||
| /// sends (BOX-071). | ||
| pub fn set_client_projection_digest(&mut self, digest: impl Into<String>) { |
There was a problem hiding this comment.
C1 · No caller in min uses this setter yet. A real create stores only projection_digest.daemon. Call it from crates/minimal/src/cmd/boxes/create.rs with the client projection digest (#3377, T128). Gatehouse §6.3.2 line 649 says the client and the node each compute the projection.
There was a problem hiding this comment.
Disputed — not addressed: The client call belongs in crates/minimal/src/cmd/boxes/create.rs, and this task's layers do not name that file. The client wiring is a residual for its own issue (#3377), as the earlier reply on this line says.
| let digests = projection_digests(&config, spec.as_ref(), verb_type); | ||
| // A client value that is not a digest in form is not stored. | ||
| if !matches!(digests.client, ClientDigest::Digest(_)) { | ||
| config | ||
| .attrs | ||
| .remove(minimald_rpc::CLIENT_PROJECTION_DIGEST_ATTR); |
There was a problem hiding this comment.
LOW · correctness · confidence 0.50 · in create_session
A client digest in form is stored even when the create carries no spec, or when the daemon computed no digest. The record then holds a client digest with no daemon digest to compare. Drop the client key in that case too, or record why it is kept.
What the reviewer checked
if !matches!(digests.client, ClientDigest::Digest(_)) { remove } -- no check on digests.daemon
Hand this to an agent
Act on a code-review finding.
File: crates/minimald/src/box_manager.rs
Lines: 4496-4501
Category: correctness
Severity: low (reviewer confidence 0.50)
Finding:
A client digest in form is stored even when the create carries no spec, or when the daemon computed no digest. The record then holds a client digest with no daemon digest to compare. Drop the client key in that case too, or record why it is kept.
First verify this against the code yourself. It is model output and
may be wrong. If it is wrong, or the tradeoff was deliberate, say so
and change nothing.
👍 useful · 👎 noise — react to tell the reviewer. Replying works too, and says more.
There was a problem hiding this comment.
Addressed in ea611e5dbebd: create_session now removes the client key when the daemon has no digest. The test asserts that a create with no spec stores neither digest.
…e-spec-and-digest-it-the-same-way-in-c
A create with no spec, or a spec the daemon cannot project, has no daemon digest. The record then drops the client key too, so it never holds a client digest with nothing to compare. The log still names it.
|
Round 9 of 10 on the runtime's review findings, pushed as
Verify Converted to a draft: something above needs a person before review. Review threads the coder judged:
A box created with a spec has a record that holds the daemon's recomputed projection digest beside the client's digest. The record stores the client's digest only when it is in digest form and a daemon digest is present. Notes from the agent: This round adds one commit, ea611e5. A create that has no daemon digest (no spec, or a spec that the daemon cannot project) now drops the client key. The test covers the no-spec case. The client wiring in crates/minimal/src/cmd/boxes/create.rs is outside this task's layers and is still a residual for its own issue. |
There was a problem hiding this comment.
arch-review: PASS_WITH_CONDITIONS · 1 condition · the full ruling
The new commit conforms. The record now stores a client digest only when the daemon also has a digest. Architecture line 21 binds the record's projection digest to its expanded spec. The span and the log still name both values. One item stays open: min does not send the client digest on a real create. Gatehouse §6.3.2 line 649 requires both sides to compute it. C1 keeps it as residual #3377.
Ruled at ea611e5 against gominimal/arch@15b603e6afb8 by claude-opus-5-5.
There was a problem hiding this comment.
minimal-review — approval withheld
Verdict COMMENT — APPROVE withheld: 12 unresolved threads — minimal-arch[bot] crates/mfile/src/projection.rs, minimal-arch[bot] crates/mfile/src/expand.rs, minimal-arch[bot] crates/mfile/src/expand.rs, minimal-arch[bot] crates/mfile/src/projection.rs:377, minimal-arch[bot] crates/mfile/src/projection.rs, …. No findings.
Panel: 3 of 3 personas ran (correctness, security, supply-chain).
T26 — Project the spec and digest it the same way in client and daemon
slice: S6
kind: task
covers: BOX-071, BOX-142
size: L
est: 460 code, 360 tests
executor: agent
difficulty: unknown ground
model: strongest
diagnostics: daemon log names both digests
observability: create span carries both digests
layers:
- projection:
- crates/mfile/src/projection.rs (new): projection and digest
- crates/mfile/src/lib.rs: mod line
- crates/mfile/src/expand.rs: secrets authorities; fills
- crates/mfile/Cargo.toml: digest dependencies
- Cargo.lock: new dependencies
test: cargo nextest run -p mfile projection_is_the_6_3_2_field_set
test: cargo nextest run -p mfile store_reference_upstream_and_inject_written_into_spec
test: cargo nextest run -p mfile expansion_fills_token_mode_quic443_steering_before_digest
test: cargo nextest run -p mfile enrolled_and_unenrolled_expansion_identical
- daemon:
- crates/minimald-rpc/src/lib.rs: carries client digest
- crates/minimald/src/box_manager.rs: stores both digests
- crates/minimald/src/box_manager/tests.rs: tests
test: cargo nextest run -p minimald projection_digests_stored_side_by_side
produces:
- crates/mfile/src/projection.rs
consumes:
- crates/mfile/src/expand.rs
proof: A created box's record holds both projection digests.
depends on:
- blocked by #2428
- blocked by #2410
Proof
When a box is created with a spec, its record attrs now hold two digests. projection_digest.client is the digest the client sent. projection_digest.daemon is the digest the daemon recomputed from the spec it received. Each is the base64url SHA-256 of the spec's §6.3.2 projection in JCS form. The box.create span and the daemon log name both. The test projection_digests_stored_side_by_side shows the two digests in the record for a matching client, a mismatched client and a client that sent none.
Notes from the agent: Design choices to review:
min box specoutput. That test is in crates/minimal/src/tests.rs, outside the layers. The design reasoning says "expansion fills". If the spec bytes must carry the filled values, a follow-up must change that test.Tests
cargo nextest run -p mfile projection_is_the_6_3_2_field_setcargo nextest run -p mfile store_reference_upstream_and_inject_written_into_speccargo nextest run -p mfile expansion_fills_token_mode_quic443_steering_before_digestcargo nextest run -p mfile enrolled_and_unenrolled_expansion_identicalcargo nextest run -p minimald projection_digests_stored_side_by_sideVerify
min task run agent-verifytimed out after 1800s, so this PR is a draft.Base
The base branch moved while the coder worked; it was merged in before the tests ran.
Residual: #3377, planned as T128 (#3387):
crates/minimal/src/cmd/boxes/create.rscomputesExpansion::projection(&resolved_type).digest()and callsBoxConfig::set_client_projection_digeston the create. Until it lands, a real create stores only the daemon digest.Closes #2430