Nexus CerebriOPEN-SOURCE PLANNING PROJECT

Slice 2: partial Phase-B canonicalization checkpoint

Date: 2026-09-27. Software 0.2.0 remains unreleased. Phase A remains complete; Phase B is not complete. Phases C, D and E are unstarted.

Source and workspace

  • Original verified Phase-B base: 2a98cf85026176590f357f2349126112baaef5a6.
  • Resumed verified base: 35eb360f132be086c8d3df9ded7433c0f8ae5fd8. The intervening 23 changed files concern the Site only; planner/dependencies/contracts are unchanged. The clean Phase-B branch was fast-forwarded, without rewriting history.
  • Branch: codex/slice-2-phase-b-canonicalization. A managed isolated checkout preserves the original checkout's unrelated maintainer Site/documentation changes.
  • Current supplied authority: NEXUS_CEREBRI_INTELLIGENCE_ARCHITECTURE_v1.2.6_CONSOLIDATED (2).md, SHA-256 007fdbab8630ecd8119215f329d7e9e36d73076de1bef885f1055c859da2edab. It closes the two previously missing byte grammars. Section 64 requires targeted independent re-qualification of §§17.3/17.5; this checkpoint does not claim that gate.
  • Execution authority remains the final Slice-2 implementation prompt plus the Phase-B kickoff. Its multi-session rule permits the nearest coherent compiling/testable checkpoint when the remaining phase cannot be completed cleanly within the session context/budget. Weekly usage was 71% at resume and 75% at the implementation checkpoint. These are observed usage values, not invented stop thresholds or a claim of exhausted quota.

Implemented boundary

  • canonical.rs: closed CanonicalValueV1, explicit tag rank, recursive numeric/UTF-8 ordering, strict imports, source set/multiset normalization, unique object keys, canonical aliases/tokens, and final JCS writer for the contract's nonnumeric JSON subset.
  • canonical_rules.rs: all 14 exact constraint variant forms; raw canonical time ranges; IANA timezone scalars; DAILY/WEEKLY patterns; typed recurrence-source conversion; closed top-level recurrence and temporal-series forms. Source-domain interval/count limits and sorted unique weekday/evidence collections reject invalid imports.
  • fingerprint.rs: explicit eight domain prefixes, exact one-placement CandidateFingerprint, HBytes/hypothesis digest, semantic hypothesis-set bytes/identity and ordered visit identity. No planner traversal, materialization or Episode capture is introduced.
  • admission_work.rs: one pure bounded typed-request serde_json 1.0.151 calculation, consumed by the existing planner gate and available for later DI/replay. No separate copied planner boolean algorithm remains. Complete B is in 1..=262144; W=B*M; admission includes W=16777216 and rejects larger W. Serialization failures expose neither an unobserved full length nor weighted bytes. Imported observations validate fixed limits and reachable measurements; validate_budget checks the supplied M separately.

Existing Phase-A wire types, RankingFeatureSet, preferred-start resolution, ordering, candidate generation, search/proof semantics, permissions, execution, CPIR, transports, dependencies and all version authorities retain their behavior. The only production change is extraction of the existing metadata admission measurement into its shared implementation.

Test evidence

Dedicated tests cover canonical ordering, malformed scalars/variants, duplicate and unordered collections, multiplicity overflow, Unicode/JCS escaping and ordering, strict rule/recurrence members, source bounds, and the externally authored §§17.3.4/17.5.5 byte examples. H1, the single-member hypothesis set, visit order and four domain-separation sanity hashes use literal architecture expectations. Candidate identity binds only source placement.

Admission tests compare 24 typed requests against independent complete serialization over six budgets and four metadata lengths. Existing planner boundedness tests independently exercise exact serialization/work limits. New tests include B=1/0, 262144/262145, W=16777216/16777217, failure states, missing/null/value, altered limits and budget mismatch.

Initial local cargo check -p cerebri-planner --locked was blocked executing the unchanged serde_core build script by Windows application control (OS 4551). The first focused test run passed three admission tests but OS 4551 blocked evaluation_canonical execution. No binary renaming, policy change, target relocation or other security workaround was used. Subsequent ordinary required gates were able to compile and execute those same paths.

Executed local gates:

  • cargo fmt --check: passed.
  • cargo check -p cerebri-planner --locked: passed on the subsequent ordinary gate run.
  • cargo clippy -p cerebri-planner --all-targets --all-features --locked -- -D warnings: passed after fixing three collapsible-if diagnostics, without suppressions.
  • cargo test -p cerebri-planner --locked: passed, 110 tests (17 new), plus doc tests. This includes the previously blocked canonical target and the unchanged boundedness, composition/ranking, graph-oracle, randomized, lifecycle and Phase-A wire tests.
  • npm run check: passed, 22 version-authority tests and repository-policy checks.
  • git diff --check: passed.

Remote CI status is recorded in the delivery handoff after pushing; no pending or blocked remote gate is represented here as passed.

Remaining Phase B and exact next action

Resume this branch and checkpoint. Do not restart Phase A or replace the current authority. Implement the closed graph/edge registry and bounded payload-aware canonical labeling first, with source-to-bytes G1/G2/G3 and rename/permutation/symmetry/state-limit tests. Then finish full PlanningRequest/BaseScenario projection and BSF (B1), identity/revision bindings, DecisionInput (D1), DecisionObservation projections (O1/O2), and their required negative/ invariance tests. Domain-prefix helpers alone do not implement these payloads or fingerprints. Canonical duration projection and graph-internal recurrence projection also remain pending.

Re-run the prescribed Phase-B gates before claiming Phase B complete. Obtain the targeted independent byte-grammar qualification separately. No architectural conflict is currently known; that pending qualification is not supplied by implementation tests.

No Codex PR Code Review or delegated review was requested or run. No Phase C implementation, capture, telemetry, learning, provider work, release, tag, version bump or baseline-debt cleanup was performed. Do not merge this partial checkpoint as completed Phase B.