@mindrian_os/cli 2.0.0-beta.3 → 2.0.0-beta.31

package/lib/core/navigation/room-context.cjs

@@ -0,0 +1,380 @@
+'use strict';
+// Phase 141-03 (RETR-01/03/04) -- getRoomContext: the 100%-local, in-process,
+// three-leg fusion that lets Larry walk the room graph mid-conversation, seeded
+// by the last ~2 conversation turns. It is the substrate behind the Capability
+// Dial's "Context Block" reach.
+//
+// Composition NOT duplication (Canon Part 7): this module copies the compose-
+// from-the-shipped-readers SHAPE of the Brain-packet builder but calls the three
+// SHIPPED legs as-is. It NEVER imports the egress projection helpers -- those
+// hash prose under the default local_summary_only mode (the RETR-03 egress
+// antipattern). Larry needs RAW prose for in-process reasoning, so this fusion
+// reuses the room-home.cjs raw safeShape path instead.
+//
+// Canon Part 8: the output feeds Larry's IN-PROCESS reasoning and MUST NOT cross
+// the wire. Zero Brain calls, zero hashing, zero non-SQLite reads (Part 9). The
+// db handle is owned by the caller (mirror evidence-claim.cjs): this module
+// NEVER requires node:sqlite and NEVER opens room.db, keeping it inside the
+// navigation allow-list (scripts/check-substrate.cjs regex /^lib\/core\/navigation\//).
+//
+// Canon Part 9: getRoomContext is the "Larry explains" face of SQL-as-local-mind.
+// SELECT supersedes folder scanning. House rule: hyphens only, no em-dashes.
+
+const { getNeighborhood } = require('./neighborhood.cjs');
+const { getRoomHomeView } = require('./room-home.cjs');
+const { getSessionHistory } = require('../memory-ops.cjs');
+// Phase 160-03 (R5 / D-03): the APP-SIDE recency decay blend. Leg D fetches the
+// cortex recency-ordered in SQL, then blends 0.995^(delta-h) here -- the visible
+// decay constant lives in recency-decay.cjs so the Leg D ranking is frozen-tested.
+// recency-decay.cjs is pure (zero requires) so it is a safe eager import.
+const { rankByRecency } = require('../temporal/recency-decay.cjs');
+// reference-now.cjs is loaded LAZILY inside resolveReferenceMs (not at module
+// init) to break a require cycle: reference-now -> feynman/timeline-renderer ->
+// navigation.cjs -> navigation/room-context (this file). A lazy require resolves
+// the module after the cycle has settled, so getReferenceNow is fully defined at
+// call time. The golden-file guard injects opts.now, bypassing this path entirely.
+
+// Leg B windowing defaults (Claude's Discretion per CONTEXT D-04 / RESEARCH
+// "Window size N"): the most recent session, its last ~6 fragments, each capped
+// so an unbounded fragment dump cannot blow the 1200ms NAV budget (T-141-05).
+const DEFAULT_FRAGMENT_WINDOW = 6;
+const DEFAULT_FRAGMENT_CHAR_CAP = 400;
+const DEFAULT_TOP_K = 10;
+const DEFAULT_MAX_DEPTH = 2;
+// How many of the most recent windowed fragments seed the focus-node resolver.
+const SEED_FRAGMENT_COUNT = 2;
+
+// Phase 270-09 (RESEARCH.md 3.2, Assumption A1): a char-count proxy, not a
+// tokenizer. RESEARCH.md's Don't Hand-Roll table forbids vendoring a
+// tokenizer for an estimate. This number is comparable only against another
+// number produced by this same divisor (the same discipline
+// tests/test-270-tool-schema-budget.cjs's approxTokens follows).
+const CHARS_PER_TOKEN_PROXY = 4;
+
+// Leg A: room-state summary via the shipped raw-prose driver. Reuse as-is.
+function legA(db, roomId, opts) {
+  try {
+    return getRoomHomeView(db, roomId, opts);
+  } catch (_) {
+    return null;
+  }
+}
+
+// Leg B: window the verbatim session history to the most recent session's last
+// N fragments, with a per-fragment char cap. RAW role/content/timestamp out --
+// never a hash. This windowing/trim is the net-new step over getSessionHistory.
+async function legB(db, fragmentWindow, charCap) {
+  let sessions = [];
+  try {
+    sessions = await getSessionHistory(db, 1);
+  } catch (_) {
+    sessions = [];
+  }
+  if (!Array.isArray(sessions) || sessions.length === 0) return [];
+
+  const mostRecent = sessions[0];
+  const frags = Array.isArray(mostRecent.fragments) ? mostRecent.fragments : [];
+  const windowed = frags.slice(-fragmentWindow);
+
+  return windowed.map((f) => {
+    const content = typeof f.content === 'string' ? f.content : '';
+    const capped = content.length > charCap ? content.slice(0, charCap - 3) + '...' : content;
+    return {
+      role: f.role,
+      content: capped, // RAW prose (RETR-03) -- never hashed.
+      timestamp: f.timestamp,
+      sectionContext: f.section_context || null,
+    };
+  });
+}
+
+// Leg C seed resolver (the load-bearing new wiring). Take the last ~2 windowed
+// fragments and resolve a focus node:
+//   1. match a fragment's section_context to a section:<context> node id
+//   2. cheap lexical fallback against nodes.properties for the same section
+// Guard with the ensureNodeExists idiom (mirror focus.cjs:101-104) before
+// returning the id. A bad/unknown seed returns null, so Leg C degrades to [].
+function resolveSeedNode(db, windowedFragments) {
+  if (!Array.isArray(windowedFragments) || windowedFragments.length === 0) return null;
+  const seeds = windowedFragments.slice(-SEED_FRAGMENT_COUNT);
+
+  // Pass 1: section_context -> section:<context> node id (the primary path).
+  for (let i = seeds.length - 1; i >= 0; i--) {
+    const sc = seeds[i] && seeds[i].sectionContext;
+    if (typeof sc === 'string' && sc.length > 0) {
+      const candidate = 'section:' + sc;
+      const row = db.prepare('SELECT id FROM nodes WHERE id = ?').get(candidate);
+      if (row) return candidate;
+    }
+  }
+
+  // Pass 2: cheap lexical fallback -- the most recently touched node whose
+  // source_section matches a seed fragment's section_context. Bound params
+  // only (T-141-06: never string-concat fragment text into SQL).
+  for (let i = seeds.length - 1; i >= 0; i--) {
+    const sc = seeds[i] && seeds[i].sectionContext;
+    if (typeof sc === 'string' && sc.length > 0) {
+      const row = db.prepare(
+        'SELECT id FROM nodes WHERE source_section = ? ORDER BY last_seen_at DESC LIMIT 1'
+      ).get(sc);
+      if (row && row.id) return row.id;
+    }
+  }
+
+  return null;
+}
+
+// Leg C: ranked graph neighbors around the conversation-derived focus node.
+function legC(db, focusNodeId, topK, maxDepth) {
+  if (!focusNodeId) return [];
+  try {
+    return getNeighborhood(db, focusNodeId, { topK, maxDepth }) || [];
+  } catch (_) {
+    return [];
+  }
+}
+
+// Phase 150-04 (MEM-03): the projected cortex node types Leg D SELECTs over.
+// These are the 150-01 writer node types (lib/core/navigation/memory-artifacts.cjs):
+//   memory_artifact   -- the 6 user-memory MD files (ROOM/STATE/MINTO/BRAIN/FEYNMAN/USER)
+//   governing_thought -- the MINTO governing-thought projection
+//   navigator_persona -- the USER role-blend x journey-stage projection
+//   decision          -- the MINTO/decisions decision_log projection (TRUTH-CLAIM)
+//   claim             -- the Phase 150.8 typed truth-claim (DIKW knowledge node)
+// Frozen so the SELECT cannot be widened by a caller; legD stays a closed read.
+//
+// Phase 150.8-04 (DIKW-09): 'claim' is added ADDITIVELY so legD surfaces typed
+// truth-claim nodes (lib/core/navigation/typed-claim.cjs writeClaimNode) to the
+// cortex-reach adapter (lib/hmi/cortex-reach-adapter.cjs CORTEX_NODE_TYPES mirrors
+// this list per the lockstep comment). legD still returns RAW-LOCAL node fields
+// only (id / type / properties / review_status) -- presence + enums, NO prose
+// egress (the Part-8 LOCAL-reasoning discipline below is unchanged).
+const CORTEX_NODE_TYPES = Object.freeze([
+  'memory_artifact',
+  'governing_thought',
+  'navigator_persona',
+  'decision',
+  'claim',
+]);
+
+// Leg D: surface the projected cortex as RAW-LOCAL node fields, RECENCY-RANKED
+// (Phase 150-04 MEM-03 + Phase 160-03 R5 / D-03).
+//
+// A single caller-owned-db, in-room SELECT over the five 150-01 projected cortex
+// node types. Mirrors legA/legC: it NEVER opens room.db, NEVER imports the egress
+// projection helpers, and returns RAW-LOCAL node fields (id / type / properties /
+// review_status + the new created_at / last_seen_at date scalars) exactly like the
+// other legs -- NO content-digest hashing, NO egress. This is the Part-8
+// LOCAL-reasoning discipline the module header mandates: the cortex feeds Larry's
+// in-process routing, never the wire (threat T-150-04-01).
+//
+// Phase 160-03 (R5 / D-03): the cortex no longer sorts alphabetically. It SELECTs
+// created_at + last_seen_at and fetches ORDER BY created_at DESC (the recency-
+// ordered fetch), then blends the 0.995^(delta-h) decay APP-SIDE via rankByRecency
+// against getReferenceNow() so the more-recent cortex node ranks first. The decay
+// constant is visible in recency-decay.cjs (never buried in this SQL), and the
+// determinism is guarded by tests/test-legd-recency-golden.cjs. The reference
+// clock is injectable via opts.now (the D-01a options.now seam) so the golden-file
+// guard fixes a frozen reference; absent an injected clock, getReferenceNow()
+// (Phase 160-01) supplies the authoritative LOCAL reference -- no network read.
+//
+// Defensive: a single try/catch defaulting to [] so a legD fault degrades
+// cortexNodes to [] without blowing the 1200ms NAV budget (threat T-150-04-04)
+// and without disturbing the other legs. The properties string is parsed
+// best-effort; a malformed properties row falls back to the raw string.
+function legD(db, _roomId, opts) {
+  if (!db || typeof db.prepare !== 'function') return [];
+  try {
+    const placeholders = CORTEX_NODE_TYPES.map(() => '?').join(', ');
+    const rows = db.prepare(
+      'SELECT id, type, properties, review_status, source_section, created_at, last_seen_at ' +
+      'FROM nodes WHERE type IN (' + placeholders + ') ' +
+      'ORDER BY created_at DESC'
+    ).all(...CORTEX_NODE_TYPES);
+    if (!Array.isArray(rows)) return [];
+    const mapped = rows.map((r) => {
+      let props = null;
+      if (typeof r.properties === 'string' && r.properties.length > 0) {
+        try { props = JSON.parse(r.properties); } catch (_) { props = r.properties; }
+      }
+      return {
+        id: r.id,                       // RAW-local node id, present verbatim.
+        type: r.type,                   // one of CORTEX_NODE_TYPES.
+        properties: props,              // RAW-local properties (handles/enums only by 150-01 contract).
+        review_status: r.review_status, // proposed/confirmed -- the truth state.
+        sourceSection: r.source_section || null,
+        created_at: typeof r.created_at === 'number' ? r.created_at : null, // RAW-local date scalar.
+        last_seen_at: typeof r.last_seen_at === 'number' ? r.last_seen_at : null, // RAW-local date scalar.
+      };
+    });
+    // APP-SIDE decay blend (D-03): resolve the authoritative reference clock once
+    // (opts.now wins for test determinism; else getReferenceNow()), then rank.
+    const referenceMs = resolveReferenceMs(opts);
+    return rankByRecency(mapped, referenceMs);
+  } catch (_) {
+    return [];
+  }
+}
+
+// Resolve the reference clock for the Leg D recency blend. opts.now is the D-01a
+// clock seam: a number, or a function returning a number (the options.now
+// closure shape), wins so the golden-file guard injects a frozen reference.
+// Absent an injected clock, getReferenceNow() supplies the authoritative LOCAL
+// reference (no network). Degrades to Date.now() if anything throws.
+function resolveReferenceMs(opts) {
+  try {
+    const now = opts && opts.now;
+    if (typeof now === 'number' && Number.isFinite(now)) return now;
+    if (typeof now === 'function') {
+      const v = now();
+      if (typeof v === 'number' && Number.isFinite(v)) return v;
+    }
+    // Lazy require breaks the reference-now -> timeline-renderer -> navigation ->
+    // room-context cycle (see the import note above).
+    const { getReferenceNow } = require('../temporal/reference-now.cjs');
+    const ref = getReferenceNow(opts && typeof opts === 'object' ? opts : undefined);
+    if (ref && typeof ref.referenceMs === 'number' && Number.isFinite(ref.referenceMs)) {
+      return ref.referenceMs;
+    }
+  } catch (_) { /* degrade below */ }
+  return Date.now();
+}
+
+/**
+ * Local in-process 3-leg fusion. 100% local. NEVER egresses. Canon Part 8 + 9.
+ *
+ * @param {import('node:sqlite').DatabaseSync} db caller-owned room.db handle
+ * @param {string} roomId the room id (for Leg A's room: root composition)
+ * @param {object} [opts] { seedFragments?, topK?, fragmentWindow?, maxDepth? }
+ * @returns {Promise<{summary, recentMessages, relevantNodes, cortexNodes, _meta}>}
+ */
+async function getRoomContext(db, roomId, opts) {
+  const options = opts || {};
+  const fragmentWindow = Number.isInteger(options.fragmentWindow) && options.fragmentWindow > 0
+    ? options.fragmentWindow
+    : DEFAULT_FRAGMENT_WINDOW;
+  const charCap = Number.isInteger(options.fragmentCharCap) && options.fragmentCharCap > 0
+    ? options.fragmentCharCap
+    : DEFAULT_FRAGMENT_CHAR_CAP;
+  const topK = Number.isInteger(options.topK) && options.topK > 0 ? options.topK : DEFAULT_TOP_K;
+  const maxDepth = Number.isInteger(options.maxDepth) && options.maxDepth >= 1
+    ? options.maxDepth
+    : DEFAULT_MAX_DEPTH;
+
+  const tA = process.hrtime.bigint();
+  const summary = legA(db, roomId, options);
+  const tB = process.hrtime.bigint();
+
+  // Leg B: prefer caller-provided seedFragments (the per-turn hot-path lane),
+  // else window the most recent stored session.
+  let recentMessages;
+  if (Array.isArray(options.seedFragments) && options.seedFragments.length > 0) {
+    recentMessages = options.seedFragments.slice(-fragmentWindow).map((f) => {
+      const content = typeof f === 'string' ? f : (f && (f.content || f.text) || '');
+      const capped = content.length > charCap ? content.slice(0, charCap - 3) + '...' : content;
+      return {
+        role: (f && f.role) || 'user',
+        content: capped,
+        timestamp: (f && f.timestamp) || null,
+        sectionContext: (f && (f.section_context || f.sectionContext)) || null,
+      };
+    });
+  } else {
+    recentMessages = await legB(db, fragmentWindow, charCap);
+  }
+  const tC = process.hrtime.bigint();
+
+  // Leg C: derive the focus node from the last ~2 messages, then rank neighbors.
+  const seedNodeId = resolveSeedNode(db, recentMessages);
+  const relevantNodes = legC(db, seedNodeId, topK, maxDepth);
+  const tD = process.hrtime.bigint();
+
+  // Leg D (Phase 150-04 MEM-03 + Phase 160-03 R5): surface the projected cortex
+  // RAW-LOCAL, recency-ranked app-side. options is threaded so the golden-file
+  // guard can inject a frozen reference clock via options.now (D-01a). Defaults
+  // to [] on any partial-leg failure, so the field is always an array.
+  const cortexNodes = legD(db, roomId, options) || [];
+  const tEnd = process.hrtime.bigint();
+
+  // Phase 270-09 (MEMOP-11/12): a char/byte cost estimate alongside the
+  // existing legTimingsMs. Cheap by construction -- the legs already have
+  // their strings in hand at this point, so this ADDS a measurement pass,
+  // never a second fetch. Measured in BYTES via Buffer.byteLength(JSON.
+  // stringify(leg), 'utf8'), not JS string .length, so a multi-byte
+  // character never silently under-counts (the same convention
+  // tests/test-270-tool-schema-budget.cjs's own measure() uses). Mirrors
+  // legTimingsMs's exact shape ({ legA, legB, legC, legD }) plus a `total`.
+  function legBytes(leg) {
+    try {
+      return Buffer.byteLength(JSON.stringify(leg === undefined ? null : leg), 'utf8');
+    } catch (_) {
+      return 0;
+    }
+  }
+  const legCostBytes = {
+    legA: legBytes(summary),
+    legB: legBytes(recentMessages),
+    legC: legBytes(relevantNodes),
+    legD: legBytes(cortexNodes),
+  };
+  legCostBytes.total = legCostBytes.legA + legCostBytes.legB + legCostBytes.legC + legCostBytes.legD;
+  const legCostChars = legCostBytes; // byte count is the measurement used; see the comment above.
+  const legCostTokensApprox = {
+    legA: Math.round(legCostBytes.legA / CHARS_PER_TOKEN_PROXY),
+    legB: Math.round(legCostBytes.legB / CHARS_PER_TOKEN_PROXY),
+    legC: Math.round(legCostBytes.legC / CHARS_PER_TOKEN_PROXY),
+    legD: Math.round(legCostBytes.legD / CHARS_PER_TOKEN_PROXY),
+    total: Math.round(legCostBytes.total / CHARS_PER_TOKEN_PROXY),
+  };
+
+  // opts.estimateOnly (MEMOP-12): "see the cost before you pay it". All four
+  // legs above are ALREADY cheap and bounded by design (the window/cap/topK/
+  // frozen-5-types discipline this file's own comments document, inside the
+  // same 1200ms NAV budget every normal call already respects) -- so the
+  // honest "cheap structural work" IS the real computation above; there is
+  // no cheaper-but-inexact path for any of the four legs that would not
+  // itself risk mis-sizing the estimate. Every leg below is therefore an
+  // EXACT size, not a guess: estimateOnly nulls the BODIES the caller would
+  // otherwise pay context-window cost to receive, while keeping the exact
+  // cost numbers already computed above.
+  const estimateOnly = options.estimateOnly === true;
+
+  // Defensive structured return even on partial-leg failure. cortexNodes is
+  // purely ADDITIVE: every existing field (summary, recentMessages, relevantNodes,
+  // _meta) stays byte-stable so the Phase 142 getRoomContext->decide() wire and the
+  // dial's existing reads are unaffected. legCostChars/legCostTokensApprox and
+  // estimateOnly are ALSO purely additive (Phase 270-09): a call with no
+  // opts.estimateOnly returns byte-identical bodies to before this change,
+  // plus the two new _meta keys.
+  return {
+    summary: estimateOnly ? null : summary,
+    recentMessages: estimateOnly ? null : recentMessages,
+    relevantNodes: estimateOnly ? null : relevantNodes,
+    cortexNodes: estimateOnly ? null : cortexNodes,
+    _meta: {
+      seedNodeId: seedNodeId || null,
+      fragmentWindow,
+      topK,
+      maxDepth,
+      legTimingsMs: {
+        legA: Number(tB - tA) / 1e6,
+        legB: Number(tC - tB) / 1e6,
+        legC: Number(tD - tC) / 1e6,
+        legD: Number(tEnd - tD) / 1e6,
+      },
+      legCostChars,
+      legCostTokensApprox,
+      estimateOnly,
+    },
+  };
+}
+
+module.exports = {
+  getRoomContext,
+  // Phase 150.8-04 (DIKW-09): surfaced so the cortex-reach claim-branch test can
+  // assert the CORTEX_NODE_TYPES mirror between this module and the adapter.
+  CORTEX_NODE_TYPES,
+  // Phase 270-09 (MEMOP-11/12): the single named home for the char/token
+  // divisor, exported so a future recalibration is one edit.
+  CHARS_PER_TOKEN_PROXY,
+};

package/lib/core/navigation/room-home.cjs

@@ -0,0 +1,142 @@
+'use strict';
+// Phase 109-09 Room Home Driver. Per CONTEXT D-08 + RESEARCH section 6.
+// Composition NOT duplication. ZERO new SQL queries; 8 reads via existing helpers + 3 raw SELECTs.
+//
+// Canon Part 1: getRoomHomeView IS the wicked navigator's working memory made legible.
+// Canon Part 4: every payload is composed from typed graph queries; the navigator never
+// re-derives; the same node id never appears twice across confirmedFacts and riskyAssumptions
+// (composition-not-duplication invariant asserted in tests/test-room-home-vs-brain-derivation-regression.cjs).
+// Canon Part 7: composes Plan 109-04 getNeighborhood + Plan 109-05 findContradictions /
+// findOpenQuestions / findRelevantOpportunities + Plan 109-03 findRecentChanges; ZERO
+// new SQL queries beyond the 3 thin raw SELECTs that scope confirmed-facts / risky-assumptions /
+// evidence-by-tier (each a single SELECT against nodes; cheaper than a CTE for small N).
+// Canon Part 9: SELECT supersedes folder scanning; this is the user-facing payoff of Phase 109.
+
+const { findContradictions, findOpenQuestions, findRelevantOpportunities } = require('./insights.cjs');
+const { findRecentChanges } = require('./memory-events.cjs');
+
+const TIERS = ['academic', 'operational', 'practitioner', 'none'];
+
+function getCurrentThesis(db) {
+  try {
+    const row = db.prepare("SELECT value FROM identity WHERE key = 'thesis'").get();
+    return row && typeof row.value === 'string' ? row.value : '';
+  } catch (_) {
+    return '';
+  }
+}
+
+function safeShape(row) {
+  let summary = '';
+  try {
+    const props = JSON.parse(row.properties || '{}');
+    summary = props.summary || props.claim || props.title || '';
+  } catch (_) { /* ignore */ }
+  return {
+    id: row.id,
+    type: row.type,
+    summary: summary.length > 120 ? summary.slice(0, 117) + '...' : summary,
+    reviewStatus: row.review_status,
+    confidence: row.confidence,
+    lastSeenAt: row.last_seen_at,
+  };
+}
+
+function getConfirmedFacts(db) {
+  // Phase 129.5-03 (TRUTH-CONFIRMED-FACTS): a human APPROVE at a Decision Gate
+  // promotes a PROPOSED truth-claim node to 'confirmed' (Canon Part 9 role 5).
+  // confirmedFacts must surface that freshly human-confirmed node, so the
+  // review_status filter is WIDENED from the 109 'validated'-only contract to
+  // IN ('confirmed','validated') and the type set is WIDENED to the full
+  // truth-claim set {claim, CausalClaim, assumption, decision, opportunity}.
+  // A confirmed decision / claim / opportunity is unambiguously a confirmed fact;
+  // a validated node still surfaces (109 back-compat). The disjointness invariant
+  // with getRiskyAssumptions is preserved by narrowing riskyAssumptions to the
+  // 'needs_evidence' state only (see getRiskyAssumptions below): a confirmed
+  // assumption cleared human review, so it belongs in confirmedFacts, not in the
+  // risky set.
+  const rows = db.prepare(
+    "SELECT id, type, properties, review_status, confidence, last_seen_at FROM nodes "
+    + "WHERE type IN ('claim','CausalClaim','assumption','decision','opportunity') "
+    + "AND review_status IN ('confirmed','validated') "
+    + "ORDER BY last_seen_at DESC LIMIT 50"
+  ).all();
+  return rows.map(safeShape);
+}
+
+function getRiskyAssumptions(db) {
+  // Per CONTEXT D-08: risky = assumption nodes still awaiting evidence.
+  // Phase 129.5-03: NARROWED from ('confirmed','needs_evidence') to
+  // ('needs_evidence') only, because getConfirmedFacts now claims confirmed
+  // truth-claim nodes (including a confirmed assumption that cleared human
+  // review). Composition-not-duplication: this set is disjoint from
+  // confirmedFacts by review_status -- confirmedFacts is the {confirmed,
+  // validated} truth-claim set; riskyAssumptions is the needs_evidence
+  // assumption set, and the two state sets do not overlap.
+  const rows = db.prepare(
+    "SELECT id, type, properties, review_status, confidence, last_seen_at FROM nodes "
+    + "WHERE type = 'assumption' AND review_status IN ('needs_evidence') "
+    + "ORDER BY confidence ASC, last_seen_at DESC LIMIT 50"
+  ).all();
+  return rows.map(safeShape);
+}
+
+function getEvidenceByTier(db) {
+  // Per CONTEXT D-08 evidence shape: { academic, operational, practitioner, none }
+  // grouped from evidence nodes by properties.tier. JS-side grouping is cheaper than
+  // GROUP BY for typical evidence counts (< 200) and preserves ordering by confidence.
+  const rows = db.prepare(
+    "SELECT id, type, properties, review_status, confidence, last_seen_at, "
+    + "json_extract(properties, '$.tier') AS tier "
+    + "FROM nodes WHERE type = 'evidence' "
+    + "ORDER BY confidence DESC LIMIT 200"
+  ).all();
+  const out = { academic: [], operational: [], practitioner: [], none: [] };
+  for (const r of rows) {
+    const tier = TIERS.includes(r.tier) ? r.tier : 'none';
+    out[tier].push(safeShape(r));
+  }
+  return out;
+}
+
+function getRoomHomeView(db, roomId, opts) {
+  const options = opts || {};
+  const roomRootId = 'room:' + roomId;
+  const since24h = Date.now() - 24 * 60 * 60 * 1000;
+
+  // 8 reads total: 1 identity + 3 raw SELECTs + 4 helper calls.
+  const currentThesis = getCurrentThesis(db);
+  const confirmedFacts = getConfirmedFacts(db);
+  const riskyAssumptions = getRiskyAssumptions(db);
+  const evidence = getEvidenceByTier(db);
+  const contradictions = findContradictions(db, roomRootId);
+  const openQuestions = findOpenQuestions(db, roomId);
+  const recentChanges = findRecentChanges(db, since24h, { limit: 20 });
+  const bankedOpportunities = findRelevantOpportunities(
+    db,
+    roomRootId,
+    { topK: 5, _mocks: options._mocks, roomDir: options.roomDir }
+  );
+
+  // Per RESEARCH section 6.1 line 882: nextMove ships a templated default in Phase 109;
+  // Phase 110 wires the live Brain advisory. The opts.brainAvailable + opts.getBrainAdvisory
+  // seam is forward-compat; until Phase 110 is shipped, the templated string is the
+  // canonical answer.
+  const nextMove = options.brainAvailable === true && typeof options.getBrainAdvisory === 'function'
+    ? options.getBrainAdvisory(roomRootId)
+    : 'See open questions and contradictions to decide next focus.';
+
+  return {
+    currentThesis,
+    confirmedFacts,
+    riskyAssumptions,
+    evidence,
+    contradictions,
+    openQuestions,
+    recentChanges,
+    bankedOpportunities,
+    nextMove,
+  };
+}
+
+module.exports = { getRoomHomeView };