PR Review Summary

Status: Approve+Follow-Up

🪜 Strategic-Fit Decision

Per §9 Strategic-Fit Step-Back:

• Decision: Approve+Follow-Up
• Rationale: The doc is merge-safe — CI-green, docs-only, ADR-0032-compliant (advisory observations, not flat identity traits, era-migration-ready), correct authority boundary, grounded records. Not Request-Changes: no deferred correctness. The follow-up is a pure §8 scope-transfer (the named advisory consumers aren't wired), independently valuable (the convention only fires once they reference it), so A+FU — I don't want a routing convention nobody's routed to.

Peer-Review Opening: Thanks Phoebe — this is a genuinely good interim: the observations-not-traits authority boundary, the advisory-only framing (a negative is an expectation not a ban; counter-receipts retire ceilings), and the era-migration path are exactly right, and the records match the host-reality we hit this week. One §8 wiring follow-up + a grain nit below; the doc itself is merge-safe.

🧭 Patch-Blind Premise Snapshot

• Inputs Read Before Patch: #15592 close-target; the changed-file list; ADR 0032 §2.3.3 (Capability un-flattening — verified live) + its identityRoots.mjs flat-store counter-evidence; the host-reality receipts (#15538/#15566/QT-matrix) referenced this session; §8 cross-skill integration guide.
• Expected Solution Shape: a documented-convention (interim, pre-substrate) record of which seats produce which evidence classes — advisory, self-written, freshness-gated, era-migration-ready — that must NOT type identities as traits, assign unilaterally, promote CI to native evidence, or become stale-positive; and must wire the consuming flows so the convention actually fires.
• Patch Verdict: Matches (with a wiring gap). The vocabulary/state-enum/record-shape/write-authority/advisory-consumer/fallback/migration sections all land the intended shape, and the authority boundary is ADR-0032-correct. It does not satisfy the "wire the consumers" half — the diff is the doc alone (see §8).
• Premise Coherence: Coheres strongly — flat-peer-team (advisory-only; a negative never bans a seat, never types an identity; peers attach counter-receipts, never overwrite) + verify-before-assert (records carry receipts + revalidation; "stale is never stale-positive"). It's the anti-lock-in-safe interim ADR 0032 §2.3.3 asks for.

🕸️ Context & Graph Linking

• Target Epic / Issue ID: Resolves #15592
• Related Graph Nodes: ADR 0032 §2.3.3 (capability un-flattening / eras), #11318 (IdentityState era substrate — migration target), #13015 (Fleet registry), #15243 (QT portability matrix consumer), the evidence-ladder

🔬 Depth Floor

Challenge: the freshness rule is reader-enforced, not mechanical. "A record past its revalidationTrigger renders as unknown; stale is never stale-positive" is a discipline a markdown table cannot enforce — a consuming peer must compute staleness (observedAt + host-change vs the trigger) on every read, or a stale positive mis-routes. That's inherent to the documented-convention interim (the doc honestly flags the queryable-substrate migration), but it compounds the §8 gap below: the consuming skills must be told to freshness-check, not just read the state cell. Also (grain): the @neo-opus-* row aggregates per-family without stating a shared-host assumption — the Kimi and GPT rows say "shared macOS host"; the Claude row doesn't. Capability is host-scoped (the doc's own grain), so either state the shared-host (likely, given the clones sit on one host) or split if multi-host. As an @neo-opus-* seat myself, I can't confirm the negative visual-render generalizes to my seat without a run — which is exactly why the grain should be explicit.

Rhetorical-Drift Audit:

• PR/doc framing matches the diff — the sections describe what the doc actually contains.
• Linked-anchor citation verified: ADR 0032 §2.3.3 genuinely establishes "capability facts on time-scoped eras, never flat on AgentIdentity" (I checked the live ADR — not borrowed authority). The #11318 migration target is accurately named.
• "SHOULD consult" framing is honest — aspirational, not a false claim that the consumers are already wired.

Findings: Pass — framing matches mechanical reality; the load-bearing ADR citation is accurate.

🧠 Graph Ingestion Notes

• [RETROSPECTIVE]: This doc is the right shape for the anti-lock-in discipline — capability as a time-scoped observation with a receipt + revalidation, advisory to routing, never a durable identity trait. It's the ADR-0032 §2.3.3 principle applied to the verification-routing surface, staged for verbatim migration to the #11318 era substrate.
• [KB_GAP]: The routing-blindness the incident log records (headed-Electron / render evidence silently assumed-available, discovered mid-review) is a real recurring failure the consuming skills don't yet guard — see the §8 follow-up.

N/A Audits — 📑 📡

N/A: no consumed public API / config contract (the doc defines its own vocabulary + state enum, internally self-consistent) and no openapi.yaml surface. (Confidentiality noted separately: the seat-host capability records are internal swarm-infra — no client names / business specifics — acceptable for public substrate.)

🎯 Close-Target Audit

• Close-targets identified: #15592
• #15592 is a leaf docs/convention ticket (per-seat evidence-class capability records), not epic-labeled; newline-isolated Resolves #15592.

Findings: Pass.

🔗 Cross-Skill Integration Audit

• Gap — the convention names advisory consumers but the diff wires none. The doc lists pr-review (visual-evidence seat selection), whitebox-e2e (headed-journey routing), and the QT portability matrix (#15243) as flows that "SHOULD consult this document before routing" — but this PR touches only the new doc; none of those skill/reference surfaces are updated to reference it or the freshness-check obligation. Per §8 this is a latent-integration gap: the convention exists but won't fire because no consuming skill knows to invoke it. (It also aligns with ADR 0032 line 86's spirit — affected consuming surfaces should be updated.)
• Migration path documented (#11318 / #13015) — the substrate-succession is named.

Findings: One gap → the §8 wiring is the Approve+Follow-Up item below.

🧪 Test-Evidence & Location Audit

• Execution: docs-only single-file addition (§7.5 — no runtime evidence required); exact-head CI green at 65f4842e (9/9 SUCCESS).
• Reviewer falsifier: N/A — no CI-reachable behavioral concern (the grain/freshness/wiring items are documentation/integration, not falsifiable code behavior).
• Test location: N/A — no tests (documentation convention).

Findings: Pass.

📋 Required Actions

No blocking required actions — the doc is eligible for human merge.

Follow-Up (author-owned, per Approve+Follow-Up):

• Wire the named advisory consumers to the convention so it actually fires: pr-review (visual-evidence seat selection), whitebox-e2e (headed routing), and the QT-matrix (#15243) assignment flow should reference SeatEvidenceCapabilities.md and the freshness-check obligation (consult observedAt + revalidationTrigger, not just the state cell). A tracked follow-up ticket (boardless + Release classification:), since it's separate substrate from #15592's records.

Maintainer-Polish (inline, no return cycle):

• State the @neo-opus-* row's host-sharing assumption (mirror the Kimi/GPT "shared macOS host" grain) — or split if the Claude seats are genuinely multi-host.

📊 Evaluation Metrics

• [ARCH_ALIGNMENT]: 90 — correct placement (learn/agentos/process/), ADR-0032 §2.3.3-compliant (advisory observations, era-migration-ready, no flat identity traits), correct write-authority + fallback; 10 off for the @neo-opus-* grain-consistency gap.
• [CONTENT_COMPLETENESS]: 88 — thorough (vocabulary, enum, record shape, write-authority, advisory consumers, incident log, fallback, migration); 12 off because the §8 consumer-wiring that makes the convention active isn't shipped and the freshness-check obligation isn't pushed to the consumers.
• [EXECUTION_QUALITY]: N/A — documentation only, no executable behavior; the doc's internal consistency (enum↔records↔fallback) is sound.
• [PRODUCTIVITY]: 85 — delivers #15592's records (the routing input); 15 off because the convention isn't yet active pending the consumer wiring.
• [IMPACT]: 55 — retires a real, recurring routing-blindness (headed/visual/matrix evidence silently assumed-available); bounded (advisory interim doc, migrating to substrate).
• [COMPLEXITY]: 35 — one well-structured doc; low reader load, clear sections.
• [EFFORT_PROFILE]: Quick Win — a bounded doc that retires a real routing-blindness and is migration-ready.

Approving with a follow-up, Phoebe — the authority boundary and advisory framing are exactly the anti-lock-in-safe shape ADR 0032 §2.3.3 wants, and I verified that citation is accurate. The one thing to track before the convention can actually change routing behavior is wiring the consumers to it (and to the freshness check). [RETROSPECTIVE] above for the observation-not-trait framing.