Aelira v0.9.7

Security

Authentication bootstrap, magic-link consumption, and open-signup provisioning are serialized. CSRF and session-cookie scope are separated, routine authentication logs exclude sensitive identifiers, and provisioning audit attribution trusts only configured proxy networks.
Analytics, evidence reports, current-compliance projections, report artifacts, provider configuration, and institution rollups are tenant-scoped. Workspace AI credentials are encrypted and resolved only for the authenticated workspace and requested purpose.
General STEM remediation is source-bound and fail closed. Typed provenance, specialist-specific verification, saved-file reverse verification, approval invalidation, and exact mixed-region composition prevent unsupported or ambiguous semantics from being published.

General STEM semantics now covers scanned equations, ordered multi-equation screenshots, vector equations, handwritten mathematics, chemical formulas, molecular structures, commutative diagrams, and mixed visuals. Each specialist has a bounded contract, purpose routing, independent verification, and explicit manual-review outcomes.
Mixed STEM composition accepts only a fully resolved typed region graph, preserves source rendering and metadata, orders the specialist outputs deterministically, verifies the serialized candidate, and binds human approval to that exact candidate.
Institution administrators can manage canonical regulatory profiles and workspace AI providers. Compliance evidence reports, current-state projections, declared institution coverage rollups, CVD metrics, durable weekly summaries, Brightspace course discovery, Blackboard signing keys, and verified CLI report retrieval extend operator and integration coverage.
Separate API and worker liveness/readiness probes, privacy-bounded health metrics, and sustained Prometheus alerts cover unavailable APIs, missing worker heartbeats, expired leases, and stalled jobs with recovery notifications.

PDF image XObject traversal rolls back safely, API keys remain usable at narrow widths, AI-provider controls stay inside their cards on phones, and Microsoft OAuth tests exercise the real HTTP seam.
Compliance deadlines come from canonical institution profiles, dashboard remediation outcomes come from durable jobs, and each document contributes only its latest verified state.
Release image security checks inspect final installed Python packages and retain bounded, expiring governance for confirmed stale lower-layer attribution.

Open-core makes no implicit AI-vendor choice. Text generation, fallback, and semantic embedding are independently opt-in; executable tenant inference resolves only the workspace's durable provider selection.
CPU- and browser-intensive work runs only in the standalone python -m src.jobs.worker service. API processes enqueue bounded jobs and remain independently responsive; Compose supplies a single-job worker default, resource limits, killable child-process execution, durable leases, and worker-specific health reporting.
POST /education/multimedia/transcribe returns an HTTP 200 asynchronous scan handle. Brightspace single-content and batch remediation return HTTP 202 job descriptors. Clients poll the authenticated status URLs for terminal outcomes and artifact references.
The unauthenticated POST /education/focus-order/analyze and POST /education/focus-order/analyze-html endpoints are removed. The independent CLI focus command and worker/scanner FocusOrder capability remain available.

Drain active work and pause intake. Back up PostgreSQL and verify the restore path, then run alembic upgrade head explicitly and confirm the repository has one head: 20260831_institution_scope.
Set and retain BYOK_ENCRYPTION_KEY before storing workspace provider credentials. Select LLM_PROVIDER, LLM_FALLBACK_PROVIDER, and EMBEDDING_PROVIDER deliberately; no provider is selected implicitly.
If Blackboard LTI is enabled, configure a matching RSA 2048+ private/public signing-key pair. Review TRUSTED_PROXY_CIDRS, session and CSRF cookie domains, Brightspace OAuth origins, shared report-artifact storage, and the worker resource and health limits before restarting services.
Deploy API and worker from the same 0.9.7 release. Confirm API readiness, a fresh worker readiness result, queue age, failed or quarantined rows, and alert recovery before resuming intake.
Update clients for the asynchronous multimedia and Brightspace response contracts and the removed unauthenticated focus-order HTTP endpoints.
Unsupported or ambiguous STEM content remains open for human review. Recognition alone never approves or publishes an artifact; persist human acceptance and allow the saved-file verifier to revalidate the exact candidate at consumption time.
Preserve every v0.9.6 operator action in CHANGELOG.md, including remediation timeout settings and deliberate handling of pre-v0.9.5 quarantined work.

#188, #189, #191, #193, and #194 harden cookie scope, durable scans, PDF rollback, provider-neutral grounding, and admin-gated provisioning.
#201, #202, #203, #204, #206, and #207 close authentication, privacy, proxy-attribution, and administrator-handoff races.
#239, #241, #242, #243, and #245 tenant-scope analytics and evidence while projecting canonical current state and deadlines.
#249, #259, #261, and #264 make regulatory and AI-provider settings durable, workspace-scoped, provider-neutral, and executable.
#265, #266, #267, #268, #269, and #270 add signed Blackboard keys, worker isolation and health, responsive key management, verified CLI reports, and durable summaries.
#271, #272, #273, #274, #275, #276, and #277 enforce configuration parity, validate local models, fill real metrics and discovery gaps, harden OAuth and image security evidence, and declare rollup coverage.

#169, #210, #211, and #246 document the fail-closed boundary and establish scanned-region detection, exact association, typed provenance, review, and approval invalidation.
#278 and #279 add verified commutative-diagram semantics.
#280 and #281 add handwritten-math suitability and verified recognition.
#282, #283, #284, and #285 add typed chemical formulas, molecular graphs, and verified PDF association.
#286, #287, #288, and #289 add ordered multi-equation and exact vector-equation association.
#290 and #291 route typed mixed regions and compose their independently verified outputs into one ordered candidate.

Each included feature PR passed the backend, lint, dependency-security, CLI, dashboard, deterministic Chromium, linux/amd64 Docker, and linux/arm64 Docker checks at its reviewed head.
Publication remains blocked until the exact merged release commit passes the same CI gates and the signed tag satisfies every protected release workflow receipt, signature, seven-file SBOM, audit, and reproducibility check.
The mandatory manual browser walkthrough is recorded separately before tagging; automated checks alone do not satisfy that release gate.
This checked-in body is consumed verbatim by .github/workflows/release.yml.