feat(#267): adopt persist v17.4.0 resolve_scores — delete the Composer client-fold + two N+1s

[emooreatx]

persist v17.4.0 shipped resolve_scores/list_scores (FSD-005 Appendix C) — the composed (subject, dimension, trust, time, state) → verdict fold now executes substrate-side as a composite op (the #329 pattern, CIRISPersist#455/#456 executor notes honored). This PR adopts both handles, deletes the Composer client-fold's substrate reads, and removes the two N+1s the 2026-07-14 demand survey flagged.

Pins

ciris-persist v17.3.0 → v17.5.0 and ciris-edge v13.0.9 → v13.1.0 in lockstep (edge v13.1.0 is the tag that re-pins persist v17.5.0; edge and server share persist types across the Engine/SqliteBackend seam, so the graph must unify on ONE persist). Applied to the root manifest (sqlite + postgres + dev-deps) and crates/ciris-lens-core. Verify stays v10.2.0 (matches edge v13.1.0's own pin).

Site 1 — src/compose_policy.rs::compose_for_key

Before: list_attestations_for(subject) fetched the WHOLE signed history; the fold (type==SCORES, trust set, staleness, group-by-(dim,subject), aggregation, licensure cap) ran client-side — plus an N+1: two has_effective_role calls per licensure:* attester in that history.

After: ONE resolve_scores(caller, filter, policy, trace=true) per the issue. The pin-once AttestationFilter carries {subject_key_id, dimension_exact, valid_at: now, lifecycle: Live, attester_filter: Explicit(pinned trust set)}; the policy is cc-4.4.2-signed-mean / cc-4.4.2-boolean-min from the dimension's CC 3.1 polarity column. Trust gating, staleness, CEG §6.1 retraction precedence, per-attester latest-wins, and the aggregation all run inside persist's .so; the verdict comes back as a ConfidenceBand + n's + the open trace, and is mapped into the existing public Composition return type.

N+1 removed: co-steward resolution now runs over the fold's head attesters only (bounded by contributor_count, never by history size) — and only for licensure:* dimensions.

Kept intact (per the issue's constraints):

• the pure Composer::compose(&[Attestation]) — the CC 4.4 normative surface tests/compose_policy.rs drives adversarially — is untouched, as are all its tests;
• TrustSet::pin_co_steward remains (consumer pins resolve first, substrate has_effective_role second);
• compose_for_key still returns Composition (no callers existed in-tree; grep confirmed).

Composition-fidelity deltas (documented in the doc comment)

• Signature gained dimension — persist's verdict is per (subject, dimension); the returned Composition carries exactly one Verdict, ALWAYS (an empty fold is now a visible Decision::Undetermined verdict rather than an absent one).
• persist owns value and the band: Verdict::value is the fold's trace aggregate; InsufficientWitnesses → Undetermined, otherwise the consumer threshold decides over the aggregate as before.
• CC 4.4.1 weighting not yet substrate-side (Frickerian low-density, CC 3.4.7 self-track-record): contributions report weight = 1.0; persist's fold marks the server-tier column-resolver refinement as its own TODO.
• Polarity coverage: persist executes signed-mean and boolean-min; Detector/Enumerated/PositiveOnly dimensions currently fold under the signed-mean default (flagged; Verdict::polarity still reports the CC 3.1 column).
• No refusal ledger: gated/untrusted/expired rows are excluded server-side and never cross the wire (no verdict-differencing) — Composition::refusals now only carries the CC 3.4.5 re-check below.
• CC 3.4.9 licensure cap kept client-side (a CCC MUST): single-source ⇒ reported confidences capped ≤ 0.5 + single_source_licensure = true; the persist-owned value/band are NOT re-capped.
• CC 3.4.5 self-emission re-check kept (CC 3.4.7 "Both checks MUST agree"): a self-emitted capacity:* head that somehow passed the substrate gate is surfaced as a SelfEmission refusal and the verdict fails closed to Undetermined.
• CC 3.1.9.3 sole-evidence screen (slashing:* on testimonial-only evidence_refs) needs the raw corpus; it runs at admission + in the pure compose tier only (flagged).
• No CC 4.4.3.9 Policy D tie-break in the substrate fold (deterministic latest-asserted/lex-id instead).

Site 2 — src/graph_config.rs

Before: list_attestations_by(node) (every row, every type) filtered client-side to SCORES && dimension == config:v1, then per surviving row a second full list_attestations_by(node) walk + a revocations_for probe (config_key_revoked) — the O(N²) revocation re-scan N+1.

After: ONE cursor-paged list_scores("", {attesting_key_id: node, subject_key_id: node, attestation_type: scores, dimension_exact: config:v1, lifecycle: Live}, cursor, 512) seek over the V106 projection. Withdrawn/superseded/recanted rows are excluded server-side; config_key_revoked is deleted. The filter is built once and reused across pages (pin-once, Appendix C.4; #[non_exhaustive]-safe construction).

Wire behavior preserved: get/set semantics, version chaining, latest-wins fold, ConfigValue::Null tombstones, and revoke-reads-as-absent are unchanged (the per-key version fold is config semantics and deliberately stays client-side). One convention note: retraction targeting now speaks the CEG §6.1 canonical envelope member references_attestation_id (which persist's lifecycle fold honors) instead of config_key_revoked's ad-hoc attested_key_id/subject_key_ids matching — nothing in production ever emitted the ad-hoc shapes; tests/graph_config.rs::recant_row was moved to the canonical member and the recant-reads-as-absent pin still passes.

Gates

• cargo build --release -j 1 — clean (16m, 0 warnings)
• cargo test --release -j 1 — 373 passed, 0 failed, 13 ignored across 41 test binaries (incl. the compose_policy, graph_config, and release_gates suites)
• cargo clippy --all-targets -j 1 — no new warnings
• cargo fmt --all — applied

The one release-gate that asserts the substrate pin (gate1_substrate_pins_at_target) was bumped to v17.5.0 / v13.1.0 alongside the manifest.

Closes #267

🤖 Generated with Claude Code

https://claude.ai/code/session_01MT2o477HcDTJJmY8GAystF

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[emooreatx]

⚠️ Do not auto-merge — needs a human CC-conformance review first. This changes the composed-verdict path on a CC 4.4-gated surface, and the implementation flags real fidelity deltas that are NOT yet substrate-side: CC 4.4.1 weighting (contributions report weight=1.0), CC 4.4.3.9 Policy D tie-break, and the Detector/Enumerated/PositiveOnly polarities — all currently folded under persist's signed-mean rather than the CC-normative rules. The pure Composer::compose (adversarial suite) is preserved, so this is safe to hold behind review. It is intentionally OFF the 0.5.117 release path (mobile gate). Recommend: confirm each delta is an acceptable interim (or gate the new read behind a flag until persist carries the full CC 4.4 semantics) before merge.

[emooreatx]

Heads-up for the CC review: persist v17.5.2 (#461) fixes list_attestations dimension_exact being a silent no-op in v17.5.0. This PR's graph_config path uses list_scores({dimension_exact: config:v1}) — on the v17.5.0 pin that filter may have returned unfiltered rows (over-matching). Re-validate this PR against v17.5.2 (which 0.5.118 pins) before merge; the dimension_exact behavior it relies on only became real in 17.5.2.