Reconcile Hamma's agent architecture guide with the live crate boundaries

[forkwright]

Finding

Hamma's public and agent-facing architecture guides describe a different crate and cryptographic/configuration topology from the code that exists.

The generated Phase A dependency blocks are now authoritative and consistent; this finding excludes them. The remaining hand-maintained orientation prose still assigns capabilities to mitos that live in dictyon or do not exist, describes an Ed25519 identity layer absent from the workspace, and directs agents to a Figment configuration cascade that is not implemented.

Verified against main 9aa9a56223afdd47d7723104cd16a5c07d5fc722.

Evidence

Crate ownership

• README.md:40-48, AGENTS.md:39-47, CLAUDE.md:34-45, and _llm/architecture.toml:22-29 describe mitos as owning Noise framing and ACL types.
• crates/mitos/src/lib.rs:1-17 still titles itself “Hamma Core” and claims Noise framing, WireGuard key wrappers, peer identity, ACL representations, and protocol constants, but exports only capability, config, keys, and types.
• Actual Noise handshake and transport framing live under crates/dictyon/src/{noise,transport,wire}. No effective ACL representation/state exists in mitos; open Build authoritative ACL policy state before advertising or activating the data plane #67 owns building that capability before data-plane activation.
• _llm/architecture.toml declares generated = false and includes histos and hamma-derp as [[architecture.crates]] with paths that do not exist. README/AGENTS correctly call them planned, but the machine-oriented architecture inventory does not distinguish live workspace members from future destinations structurally.

Identity model

• CLAUDE.md:70 says node identity is Ed25519, WireGuard tunnel keys are Curve25519, and Noise uses X25519, with each represented by a distinct newtype.
• crates/mitos/src/keys.rs:1-12,205-237 says all current key families are Curve25519/X25519 wrappers. NodePrivate is explicitly the WireGuard identity; no Ed25519 dependency or identity type exists anywhere in the workspace.

Configuration model

• CLAUDE.md:71 says TOML configuration is parsed through Figment with an environment-variable override cascade.
• Root Cargo.toml contains no figment dependency, and a workspace source inventory finds no Figment use or TOML/env configuration loader. mitos::Config is currently a typed/default/validation structure passed directly by callers.

#118 owns the separate public Phase A order and Kanon-consumption contract. This issue does not reopen or duplicate that generated block.

Why this matters

These files are the dispatch surface for agents. Following them can lead an implementation to:

• add framing logic to the wrong crate;
• assume an ACL substrate exists when Build authoritative ACL policy state before advertising or activating the data plane #67 says it does not;
• design around a nonexistent Ed25519 identity authority;
• search for or extend a configuration cascade that has never been built; or
• treat planned crates as current dependency owners.

That is more than stale prose. It creates competing architecture authorities around security-critical boundaries where type identity and crate ownership are intended to prevent accidental mixing.

Desired correction

Rebuild the hand-maintained architecture guidance from the live workspace and explicitly separate three states:

1. implemented ownership;
2. approved but blocked/planned capability; and
3. speculative future destination.

Make mitos documentation name only the vocabulary it actually owns today. Route Noise framing to dictyon; link ACL claims to #67 until the types and state exist. Correct the identity description to the implemented Curve25519/X25519 hierarchy, or add a separately justified Ed25519 authority only when a real consumer requires it. Describe configuration as direct typed config unless/until a real loader lands.

Derive workspace-member/path inventory mechanically while retaining human-owned role descriptions.

Done when:

• every live crate is represented exactly once and planned crates cannot appear as existing paths;
• mitos docs and exports agree on its current responsibility;
• Noise, ACL, identity, and configuration guidance matches executable ownership;
• no guide claims an Ed25519 or Figment surface absent from dependencies/source;
• adding/removing/renaming a workspace crate makes the architecture inventory check fail; and
• the Phase A generated contract remains unchanged and authoritative for sequencing.