ADR 0034: ADR Candidate Registry as Architectural Memory

Status

Accepted (2026-04-26). Back-fills the artifact created in this PR. The registry exists; the ADR records the decision to keep it.

Context

ICN's architecture spans many domains: identity, governance, federation, economics, compute, member shell, accessibility, conflict resolution, package model, website truth, project governance. Decisions are recorded in docs/adr/; the lifecycle for those decisions is governed by ADR-0018.

What ADR-0018 does not address is the pre-decision layer — the long arc of architectural intent that has not yet been written as an ADR but that the project needs to remember. Without a coordination artifact for that intent, the long arc tends to fade between sessions: contributors and agents re-discover it ad hoc, and pieces get accidentally designed out.

This PR introduces ops/coordination/adr_candidates.yaml and the companion docs/strategy/ICN_CONSTITUTIONAL_ROADMAP.md as that coordination artifact. This ADR records the decision to maintain them.

Decision

ICN maintains a machine-readable ADR candidate registry at ops/coordination/adr_candidates.yaml and a human-readable companion at docs/strategy/ICN_CONSTITUTIONAL_ROADMAP.md.

Registry properties

Each candidate row describes an ADR ICN intends to write, including its target id, title, proposed status (proposed | accepted), the domain it addresses, why it matters, runtime surface, accessibility implications, conflict path, compute boundary, website truth implication, existing issues, missing issues, dependencies, priority horizon, and tranche.
Registry state progresses candidate → drafted → merged. Drafted rows reference their adr_path. Merged rows are pruned periodically as cleanup.
The registry is not authoritative. It is a planning index. ADR-0018 governs lifecycle once a row is drafted as an ADR file.
Tranche numbering groups ADRs by drafting order: tranche 1 (foundational, this PR), tranche 2 (federation/economics), tranche 3 (member shell/accessibility/conflict), tranche 4 (project governance/release/kernel-dispatch detail).

Roadmap properties

Eleven domains match the registry's domain taxonomy.
Each domain section answers four questions: why it matters, ADR candidates that belong here, existing issues, priority horizon.
Non-doctrine by explicit declaration. Decisions live in docs/adr/.
Modifiable. Forward-direction items can be modified, refused, or superseded once written as ADRs.

Maintenance

A row flips from candidate to drafted when the ADR is written and merged. The roadmap section is updated to link the ADR.
Rows that are decided to be unnecessary are removed (not silently kept as zombies).
New rows are added as new architectural concerns emerge; the registry is not closed.

Consequences

The long arc of design becomes legible. Contributors and agents can pick up tranches without rediscovery cost.
Work prioritization is visible. "What's next?" has a written answer at the architectural layer.
Trade-off: maintenance cost. The registry must be kept current or it becomes worse than nothing.
Trade-off: the registry could be misread as binding. ADR-0018's lifecycle is the actual binding mechanism; the registry is a coordination artifact only.

Implementation status

Implemented in this PR.

Evidence:

ops/coordination/adr_candidates.yaml — 60+ candidate rows across 11 domains.
ops/coordination/README.md — conventions and limits.
docs/strategy/ICN_CONSTITUTIONAL_ROADMAP.md — human-readable companion.
Tranche 1 ADRs (0021–0034) drafted concurrently with the registry.

Alternatives Considered

Alternative	Why rejected
Keep architectural intent in scattered design docs	What we had through April 2026. Scattered intent fades; long arcs get re-discovered.
Use issues alone as the planning surface	Issues are work units; they fragment the architectural arc. A row for "CCL institutional process language" is one architectural concern; the issues that fall out of it are many.
Make the registry an ADR rather than a separate artifact	The registry is meta to ADRs. Making it an ADR would mean every row update is an ADR amendment — too heavy for a coordination artifact.
Embed the registry in the roadmap document only	Machine-readable separation lets future tooling (a "next tranche" suggestor) consume it. The roadmap is the human view; the YAML is the machine view.