v1.3.0 — the epistemic-loops release: NLI adjudication, bitemporal beliefs, dream score fix

The epistemic-loops release

A deep review concluded that the cognitive architecture was sound but its self-correction loops were severed at the joints: signals written to a table nothing read, contradictions recorded but never verified, the dream cycle rating its own maintenance as evidence of use, and evidence discarded at merge points. This release closes those loops.

Added

• First-class signal reads — CortexStore.getSignal(id), getSignals(filters), updateSignal(id, updates) implemented in SQLite, Firestore, JSON, and ScopedStore. Fixes a split-brain bug where putSignal wrote a dedicated table on SQLite/JSON that surface/resolve never read — every SURPRISE and hindsight TENSION signal was invisible on the default backend. Legacy signals written through the generic collection API remain readable and updatable. Signal gains resolved_at and observation_id.
• Contradiction adjudication (engines/adjudicate.ts, docs/nli-adjudication.md) — contradict now verifies the claimed conflict before recording it: NLI cross-encoder first (both directions), LLM fallback, graceful degradation to unverified tension. Genuine contradictions apply a confidence penalty scaled by adjudicator confidence; complementary/unrelated evidence records nothing. force: true preserves caller authority. Low-tier adjudicators (per model_provenance.confidence_tiers, resolved via the new resolveModelTier()) cannot declare genuine contradictions below 0.8 confidence — the first real consumer of capability tiers.
• Observe-time implicit-conflict detection — when an NLI provider is configured, observe() adjudicates observations that land in the merge/link band against the nearest memory before the gate can reinforce the memory they dispute (negations embed close to their affirmations). Genuine → action: "contradiction" + signal + penalty; succession → action: "superseded" + revision guidance. Opt out per call with check_conflict: false.
• Bitemporal belief entries — BeliefEntry.valid_from/valid_to (valid time) alongside changed_at (system time); believe accepts valid_from. New supersedes adjudication verdict distinguishes temporal succession ("moved to Berlin" vs "lives in Paris") from same-time contradiction: succession recommends revision and applies no penalty, because a superseded belief was not wrong.
• NLI service bootstrap (scripts/nli-service/, shipped in the npm package) — FastAPI cross-encoder service matching the LocalNLIProvider wire contract. fozikio nli runs it from an installed package: first run provisions a virtualenv at ~/.fozikio/nli-venv and installs requirements, later runs start immediately (--port/--host/--model/--venv/--reinstall). Enable in the engine via nli.enabled config or CORTEX_NLI_URL.
• Versioned prompt registry (engines/prompts.ts) — every cognitive prompt in the engine (dream phases, HyDE, adjudication, salience scoring, reflect, abstract, query-explain, agent-invoke findings, ruminate) as typed, versioned templates with a pinned snapshot test forcing deliberate version bumps. Migrating ruminate also removes a latent String.replace bug where $-patterns or a literal placeholder in the gathered context could corrupt the prompt.
• Structural thought-quality gate (engines/thought-quality.ts) — dream refine/abstract output is judged by grounding (keyword overlap with the evidence it derives from) plus form checks; the empirical "foreign thought" marker lists survive as a weak corroborating signal instead of a model-specific veto.

Changed

• Dream score phase no longer counts as access. Passive FSRS review previously called touchMemory, refreshing last_accessed/updated_at — the next cycle then read its own touch as "recently accessed → rating Good", so cron dreams reinforced every scored memory without real use (the exact silent hardening the hindsight phase audits for). Scoring now writes FSRS state only; access strength comes from genuine retrieval.
• FSRS contradiction penalty now sees signal-based contradictions, not just contradicts edges — observation-vs-memory conflicts are signals (observations are not graph nodes) and were previously invisible to scoring.
• observe merge no longer discards content — merged observations are stored unprocessed so the next dream cycle clusters them into the same memory and feeds the refine phase, converting duplicates into consolidation evidence.
• Pairwise edge discovery pre-filters by embedding similarity (cosine < 0.2 skipped), cutting most of the O(n²) LLM calls without touching contradiction-range pairs.

Fixed

• npm test now works on Windows (vitest invoked via its .mjs entry instead of the POSIX bin shim); file-backed SQLite tests close handles before temp-dir cleanup (Windows EBUSY).