docs(adr): ADR-007 Rev 7 (Rule 8 principal-scoped backends) + ADR-063 comm principal model

[ohdearquant]

ADR-007 Rev 7 (Rule 8: principal-scoped pack backends) + ADR-063 (comm principal model)

Docs-only. Two ADRs, one contract addition plus its first consumer spec.

ADR-007 Rev 7 — Rule 8 (additive carve-out)

A pack MAY declare a backend whose correct trust model is principal-scoped isolation enforced server-side, rather than the shared-store attribution model that Rules 0–7 describe. Rule 8 states this carve-out precisely:

• The shared local substrate is unaffected — Rules 0–7 (including the Rev 6 episodic Rule 0 amendment) apply in full to KG, memory, gtd, brain, schedule, knowledge, and any pack on the shared SQLite backend. No post-fetch namespace check is added to by-ID ops; the Gate stays the single enforcement seam.
• A pack with a principal-scoped backend carries its own isolation contract (per ADR-028 per-pack backends). "Namespace is attribution, not isolation" remains correct for the shared store; "principal identity is the storage boundary" is correct for a remote broker. These are statements about different backends — not a contradiction.

ADR-063 (Proposed) — comm pack principal model

Specifies the comm pack as the first consumer of the Rule 8 carve-out:

• Today (OSS, single machine): actor-addressed delivery on the shared "local" store (ADR-057) — correct, with the to_actor filter dormant for the default deployment until issue design: actor identity on every request (authenticated caller, verb-level scoping) #75 lands.
• Cross-machine: lambda-to-lambda coordination needs a transport that is inherently multi-principal (each lambda authenticates independently, reads only its own inbox). A remote broker enforces per-principal scope server-side — not served by a shared SQLite store with a view-layer filter. ADR-063 specifies that contract, authorized by Rule 8.

Sibling to ADR-056 (human out-of-band channel transport); related to #75, #112, #113.

Revision-numbering note

This was drafted as Rev 5 before the episodic carve-out (Rev 6) merged to main (#177). Rev 5 never landed, so this is re-seated as Rev 7 on top of Rev 6. The two amendments are independent: Rev 6 amends the Rev 4 Rule 0 write-default for episodic memory; Rev 7 adds Rule 8 for principal-scoped backends. Neither touches the other. Rule 8's normative text is byte-identical to the original draft — only revision labels and the supersession framing were updated to sit on Rev 6.

Status

Both ADRs are design/contract — no code in this PR. Rule 8 is Accepted (additive carve-out); ADR-063 is Proposed (implementation follows separately, per ADR-driven development).

🤖 Generated with Claude Code

[ohdearquant]

Holding for Ocean: this is an architectural decision, not a process gate. The ADR pair is coherent. ADR-007 Rule 8 (Rev 7) is an additive carve-out for principal-scoped backends, e.g. a remote message broker whose server-side scope is itself a Gate implementation; it explicitly does NOT add per-record namespace checks to the shared SQLite substrate, so it does not conflict with the Rev 3+ attribution-only model. The open item is ADR-063 (comm principal model), which is Proposed with two unresolved design questions: OQ-1 credential mechanism and OQ-2 per-principal partition strategy. It commits to a remote broker for cross-machine lambda delivery, which is a strategic call and sits in tension with the current DEPTH-not-BREADTH direction. ADR-063 Step 1 (thread actor identity) also overlaps with the comm fix just merged in #203. CI is currently failing (macOS + Ubuntu) and must clear before any merge. Marking draft and leaving open for Ocean to set direction and answer OQ-1/OQ-2. (lambda:khive triage, 2026-06-21)