[shreyanshjain7174]

Plan

Phase 2 of the AgentGate scale-out prompt, this pass: ADRs only, no code. Each ADR is Proposed and needs owner approval before any implementation starts.

Files

• docs/design/README.md: index plus an Open decisions table with recommended defaults.
• docs/design/0000-phase2-sequencing.md: dependency order and what can proceed independently.
• 0001-per-replica-chains.md:
  • chain_id = <replica-id>:<epoch> as a v2 field;
  • fencing via an external lease and token, with a cached self-fence deadline;
  • an append-only replica registry, so a missing chain is reported lost;
  • migration descriptors co-signed by root and the final v1 key.
• 0002-two-phase-receipts.md:
  • intent before dispatch, outcome after;
  • outcome_unknown on crash recovery, not_dispatched only when certain at runtime;
  • upstream attempt count recorded;
  • retry dedup deferred.
• 0003-durable-shipping.md: Postgres stays parked.
  • Signed sealed SQLite segments go to S3, MinIO or a filesystem export, written with conditional puts or O_EXCL.
  • A named signer for every artifact.
  • An epoch key is destroyed only after its final descriptor is durable.
• 0004-merkle-checkpoints.md: signed head vectors and inventories first, Tessera and C2SP tlog later. A singleton fenced publisher with defined key custody.
• 0005-root-certified-keys.md:
  • an offline root with a scoped in-cluster intermediate;
  • per-epoch certs with sequence-bounded validity, where time windows are only hygiene;
  • a trust_status cutoff on chain loss or takeover;
  • rotation at a batch boundary without draining;
  • v1-style binary preimages.
• 0006-receipt-failure-modes.md: fail_closed (503 plus control receipts) ships. fail_open_spool is deferred until a concrete failure class justifies it. The deferred design keeps evidence-hiding defenses.
• 0007-kubernetes-lifecycle.md:
  • StatefulSet with one PVC per replica;
  • an empty PVC must be explicitly enrolled;
  • storage fencing is proven before takeover, with force-detach tests;
  • single replica until per-replica auth and vault state is solved;
  • a 40s drain.
• docs/spec/RECEIPT-v2.md: a proposal outline, marked "Proposed. Not implemented." v1 verification is unchanged.

Decisions for the owner

See docs/design/README.md#open-decisions. The main ones:

• trust structure and key custody;
• fencing primitive;
• registry storage backend;
• storage model and retention;
• publication (head vectors vs Tessera);
• spool;
• the latency gate. Phase 0 p99 already exceeds 1 ms under concurrency, per perf(bench): end-to-end overhead harness and Phase 0 gate results #50.

Validation

• go build ./...: passed (docs only)
• 125 relative links and anchors resolve
• No em dashes, whitespace clean

Review

Two rounds by a second model (Claude Opus 5.5). Round 1 had blocking issues:

• 0001, 0007: fencing and lost-chain detection;
• 0005: an air-gap-unusable root ceremony and time-based validity;
• 0003: undefined signers;
• 0006: a spool with little real coverage.

All are resolved. Round 2: acceptable as proposed. Its 4 non-blocking refinements are applied.