Skip to content

P.O.W.E.R. 3.5.0 release notes

This document describes the published v3.5.0 release. The signed tag, package and SBOM artifacts, release receipt, and remote readback passed the final release workflow on 2026-08-12.

Core outcome

  • Lean base installation and offline FTS are the canonical path.
  • Dense, reranker, graph, query-expansion, and ROT implementations live under power_framework.experimental; core keeps lazy compatibility shims and does not import optional adapters during startup.
  • auto reports actual_mode, fallback_reason, and index age; explicit semantic retrieval remains fail-closed when its optional profile is absent.
  • CLI, local MCP, and library callers share typed application envelopes for discovery, retrieval, proposal/apply, receipts, handoff, and degraded fleet status.
  • Portable lifecycle adapters cover session-start, post-write verification, pre-compact checkpoint proposals, and stop/drift checks without silent semantic writes; the fresh-agent chaos suite records eight redacted scenarios.
  • Doctor and bootstrap are bounded/read-only by default; proposal content is accepted through a file or stdin rather than process arguments.
  • Maintenance is preview-first, exact-hash-bound, reversible for safe_auto, and plan-only for retention/deletion candidates.
  • The retention soak proves exact preview/applied manifests, bounded backup growth, newest-valid-backup restore, and preservation of source/control files.
  • The cheap health loop deduplicates issue notifications with exponential backoff; opt-in external evidence captures exact bytes in a content-addressed store and fails closed on tamper or unavailable-source readback.
  • power control-plane materializes a conflict-safe Markdown cockpit and an optional, marker-owned POWER Control.base with four views, while power migrate-state provides a content-free state-plane inventory only.
  • The synthetic Phase 8 benchmark covers 20 technical workflows and compares POWER with a declared no-POWER handoff baseline, bilingual strata, five false-premise abstentions, and a historical/current state filter; it remains technical-only and does not replace blind human or real-vault outcome evidence.
  • A separate continuity receipt runs 20 handoffs across independent worker processes and checks replay idempotency, proof-carrying resume, blocked decisions, and source preservation; it remains synthetic evidence.
  • The executable upgrade matrix also kills a subprocess at before_move, after_move, and after_pointer, then proves active-generation readback, restart recovery, source-byte preservation, and no data loss. This is a hermetic Ubuntu release matrix, not physical 3.4.5 runtime or Windows/macOS compatibility evidence; macOS and Windows are deferred from this release.
  • Crash recovery reaps only lease-expired building rows and their UUID-owned staging/generation artifacts; fresh rows are retained for live writers. CI uploads one content-free Ubuntu report and blocks on its aggregate.

Evidence boundary

The release gate records local tests, lint, types, documentation drift, the FTS-first packaging contract, and the aggregated Ubuntu executable upgrade matrix. macOS and Windows are explicitly deferred and are not supported-platform claims for v3.5.0. Dense quality, real-vault outcome improvement, blind sealed-human scoring, and remote release readback were separate mandatory evidence gates. Physical Windows/macOS migration remains deferred and is not a v3.5.0 supported platform claim. The final workflow received the two content-free Phase 8 receipts described in release/evidence/README.md through the protected power35-stable-release environment; it did not infer them from a passing unit suite.

Validation commands

./.venv/bin/pytest
./.venv/bin/ruff check src tests scripts
./.venv/bin/ruff format --check src tests scripts
./.venv/bin/mypy src/power_framework
./.venv/bin/python scripts/check_doc_drift.py
./.venv/bin/python scripts/verify_upgrade_matrix.py --from-version 3.4.5 --to-version 3.5.0
./.venv/bin/python scripts/retention_soak.py --output /tmp/power35-retention.json

CI executes the content-free upgrade invariants on Ubuntu only for this release; macOS and Windows are deferred. The publication workflow required a clean tag-bound baseline generated by scripts/generate_release_baseline.py --tag v3.5.0 --validation-report <power.release-validation.v1.json>, validation with scripts/verify_release_contract.py --require-tag, package smoke, SBOM, and readback of the release body/assets. The resulting receipt is attached to the published GitHub release.