CortexDB 0.9.3 — GDPR erasure integrity, audit hash chain, temporal phrases (connectors 0.2.14)

Bug-fix and hardening release driven by the KAN board triage: 129 open issues were investigated against the codebase, ~75 verified server and connector defects were fixed, and every ticket got an evidence-backed disposition. No storage format changes, no migration: 0.9.x data directories work as-is (new audit-chain and vocabulary fields are serde-defaulted; legacy records replay cleanly).

GDPR / erasure integrity (the headline)

• forget cascade=derived_only actually cascades (KAN-9/28/33/37/67): callers sending layers:["events"] previously zeroed out the derived cascade set, leaving episodes/facts/beliefs/understanding containing erased text. Derived layers explicitly named still narrow the cascade; when none is named, all four derived layers are cascaded.
• Erasure previews honor the selector (KAN-6/30/38/69): /v1/erasures/preview accepted only {scope} and estimated the whole scope for deletion regardless of any selector. It now takes the same selector as execute (deny_unknown_fields), dry-runs the real derived match, and returns a per-layer estimated_affected breakdown.
• Two-phase forget (KAN-12): new non-destructive POST /v1/forget/preview returns a fpvw_ preview id + per-layer estimated_deleted; /v1/forget validates from_preview_id.
• Erased data cannot resurface via time-travel (KAN-11/70/71): the recall engine drops episodes/concepts whose cited source events no longer resolve — unconditionally, and before temporal gating, so as_of pins can't serve pre-erasure snapshots of erased content.
• Event-derived concepts are reachable by the cascade (KAN-10): SupportedBy gained an events provenance field, populated by the concept reflector and matched by the deletion cascade.
• Reference scrubbing covers labels (KAN-32): surviving events lose supports:<erased_id> labels, not just preceded_by references.

Audit: tamper-evidence is now real

• Request bodies are actually hashed (KAN-41): the audit middleware hashed empty bytes for every request, making /v1/audit/verify vacuous. It now buffers and hashes the raw wire bytes (multipart, oversize, and streaming bodies get explicit unhashed: markers). POST /v1/forget appends a real gdpr=true audit row (was minting a throwaway id).
• Hash-chained rows (KAN-26/39/68): row_hash = sha256(prev_hash || canonical_json(row)) computed on append; new GET /v1/audit/chain (head hash, length) and POST /v1/audit/chain/verify (reports first_broken_index on tamper, deletion, or splice). Legacy JSONL rows are tolerated as a prefix.
• Object bodies verify (KAN-72): /v1/audit/verify accepts a raw string (hashed byte-for-byte) or a JSON object (canonical form: byte-sorted keys, no whitespace).

Temporal: the phrases people actually type

temporal.natural (KAN-7/36/54/55/62/63/65/66) now parses: bare last week/month/year, N <unit> ago including word numbers (two weeks ago, a month ago), last/this quarter, sprint phrases (two sprints ago, documented 14-day convention), early/mid/late period thirds (early last week, including with a leading noun phrase), and everything cortex-bitemporal's date grounder handles (seasons, written months, bare years) via fallback. Bi-temporal correctness fixes ride along: episode recorded-axis pins gate on recorded_from (not started_at — late-delivery episodes no longer leak into as_of snapshots, KAN-43), and backdated/imported events are visible at historical pins via source_recorded_at (KAN-27/35).

New / completed API surfaces

• Import (KAN-25/58/59/64): generic POST /v1/import; export→import roundtrips work (import falls back to the MemoryEventV1 export shape, preserving bi-temporal history via source_recorded_at); import jobs emit import_progress/import_error/import_complete lifecycle events; the 202 carries accepted/failed/first_error/status_url.
• Beliefs (KAN-31/42/44): POST /v1/beliefs creates caller-declared beliefs with event/fact supports; /v1/beliefs/build returns full diagnostics instead of a bare count; belief-shaped events materialize deterministically without LLM triples; superseded beliefs close (valid_to + Deprecated) and are filtered from recall.
• Vocabularies (KAN-8/34/51/61): terms with definition/aliases/active/valid_from/valid_to/replaced_by persist and round-trip; unknown fields 422 instead of silently dropping; PATCH /v1/vocabularies/{name}; per-tenant isolation (closes an unfiled cross-tenant overwrite hole); schema published at /v1/schemas/vocabulary.

Server hardening

• Container auth hole closed (KAN-74/85): CORTEX_BIND_ALL=1 (set in the Docker image) defeated the loopback check, so a plain docker run served 0.0.0.0 with auth disabled. Loopback detection now derives from the resolved bind address, fails secure, and keyless container boots generate a key. Startup logs report bind/loopback/auth posture.
• Scope-creator lockout fixed (KAN-60): POST /v1/scopes with an empty roster seeded a scope its own creator couldn't read. The creator is now seeded as Owner; owner-less rosters are rejected.
• --help/--version (KAN-76): short-circuit with zero side effects (previously booted the full server and created data directories).
• No unbounded waits (KAN-40): the legacy coalescer path's ack wait is capped at 30s like the unified path.

Connectors 0.2.14

• Discord: text-less messages (stickers, polls, untitled embeds, system messages) are rendered or skipped instead of being sent as empty envelopes the server permanently 422s (20 tickets, one cause).
• Confluence: pages are enumerated via the read-your-writes content REST API instead of the eventually-consistent CQL search index, so just-created pages are found in their own sync run (9 tickets).
• Notion: comments on archived parent pages are ingested; comment fetch failures are logged instead of swallowed (KAN-24).
• Google Workspace: subject-less, bodyless Gmail messages are skipped client-side.
• Teams: docstring no longer overclaims transcripts/file shares (KAN-77).

Verification

Workspace cargo check clean; ~2,050 Rust tests and 245 connector tests passing, including new regression tests pinning every fixed repro shape (the layers:["events"] cascade gate, the preview selector, the middleware body hash, the episode recorded-axis pin, the archived-parent Notion comment, and the Discord content policy among them).