Council v3: seat-local smart evidence acquisition

[Kitahl]

Scope

Implements Council v3 as an additive layer over the completed Council vNext branch.

• preserve challenge-derived seats, commit-reveal, cross-critique, overlap diagnostics, matched DIRECT control, and REVIEW-only authority
• add frozen caller-supplied total and per-seat evidence budgets
• freeze one utility/risk policy per seat outside candidate bundles
• accept neutral receipt-bound evidence bundle candidates without importing foil_*
• require task-only, dependency-complete, acyclic, read-only, capability-matched bundles
• exclude hidden-gold-dependent and over-budget candidates
• prune dominated bundles and select only positive conservative-utility bundles
• freeze all seat plans before acquisition and complete all acquisition before any first-pass commit
• forbid cross-seat result reuse and budget transfer before reveal
• prevent INVALID/UNRESOLVED tool evidence from becoming admitted evidence
• bind seat submissions to frozen baseline + admitted evidence partitions
• measure shared source/tool overlap without inferring independence
• bind a v3 evidence-audit hash into the ordinary Council REVIEW receipt

FOIL boundary

The Smart-Tool update is proposed/unvalidated and no smart-tool implementation exists in the repository yet. Council v3 therefore does not depend on or import FOIL internals. A future FOIL controller may supply neutral EvidenceBundleCandidate rows only; Council retains its frozen utility policy, evidence admission, review authority, and target-module handoff.

Authority

• evidence acquisition is READ_ONLY and EVIDENCE_ACQUISITION_ONLY
• Council still clears only REVIEW
• Council cannot clear target domain obligations, execute repairs, or authorize writes
• target modules remain responsible for claim-native verification
• no efficacy claim is made

Validation

The exact Council v3 tree 5a606e97c09e44493422e9d6b9cf1f24ee94b862 was validated through tree-identical CI trigger commit 13e6eac226fde531026697a48b0039dd024954bd, then the branch was restored to the clean one-commit implementation 7fcfd259a6b103a464061c1d30f2a48fa1ff7f52.

All required checks passed:

• Research software validation: PASS
• complete repository suite: 510 tests passed
• focused Council v3 suite: 10 tests passed
• vNext typed-runtime validator: 40/40 passed
• public showcase validator: 33/33 passed
• lint and Python compilation: PASS
• Runtime portability: PASS on Ubuntu, macOS, and Windows
• Security gates: PASS, including full-history secret scan, dependency audits, deterministic hash-lock regeneration, and exact locked installation
• CodeQL: PASS

Focused Council v3 coverage

tests/test_council_v3_evidence.py covers:

• deterministic positive-utility bundle selection
• stand-down on non-positive utility
• hidden-gold and non-read-only exclusion
• frozen total/per-seat budget conservation
• dependency-cycle rejection
• all evidence acquisition before any first-pass commit
• cross-seat result-reuse rejection
• invalid/unresolved evidence admission rejection
• frozen submission/evidence-partition binding
• source overlap without an independence claim
• REVIEW-only finalization with v3 evidence-audit binding

Base: upgrade/challenge-vnext-council at 132172944c07de77c11266c953996fc1e840b647.