diff --git a/src/headless-status.ts b/src/headless-status.ts
new file mode 100644
index 000000000..901548517
--- /dev/null
+++ b/src/headless-status.ts
@@ -0,0 +1,108 @@
+/**
+ * headless-status.ts — direct `sf headless status` implementation.
+ *
+ * Purpose: keep the headless status machine surface read-only and
+ * TTY-independent instead of routing through the interactive `/status` overlay
+ * command or the long-lived RPC/v2 session handshake.
+ */
+import { buildQuerySnapshot, type QuerySnapshot } from "./headless-query.js";
+
+export interface HeadlessStatusResult {
+	exitCode: number;
+	data?: QuerySnapshot;
+}
+
+/**
+ * Render a compact text status from the query snapshot.
+ *
+ * Purpose: provide the same operator value as `/status` in terminals where no
+ * interactive overlay can be displayed.
+ *
+ * Consumer: handleHeadlessStatus for text-mode `sf headless status`.
+ */
+export function formatHeadlessStatus(snapshot: QuerySnapshot): string {
+	const { next, runtime, uokDiagnostics, schedule } = snapshot;
+	const state = snapshot.state as any;
+	const lines = ["SF Status", ""];
+	lines.push(`Phase: ${state.phase}`);
+	if (state.activeMilestone) {
+		lines.push(
+			`Active milestone: ${state.activeMilestone.id} - ${state.activeMilestone.title}`,
+		);
+	}
+	if (state.activeSlice) {
+		lines.push(
+			`Active slice: ${state.activeSlice.id} - ${state.activeSlice.title}`,
+		);
+	}
+	if (state.activeTask) {
+		lines.push(
+			`Active task: ${state.activeTask.id} - ${state.activeTask.title}`,
+		);
+	}
+	const progress = state.progress;
+	if (progress) {
+		const parts = [
+			`milestones ${progress.milestones.done}/${progress.milestones.total}`,
+		];
+		if (progress.slices)
+			parts.push(`slices ${progress.slices.done}/${progress.slices.total}`);
+		if (progress.tasks)
+			parts.push(`tasks ${progress.tasks.done}/${progress.tasks.total}`);
+		lines.push(`Progress: ${parts.join(", ")}`);
+	}
+	if (state.nextAction) lines.push(`Next: ${state.nextAction}`);
+	if (state.blockers.length > 0)
+		lines.push(`Blockers: ${state.blockers.join("; ")}`);
+	lines.push("");
+	lines.push(
+		`Dispatch: ${next.action}${next.unitType ? ` ${next.unitType}` : ""}${next.unitId ? ` ${next.unitId}` : ""}${next.reason ? ` - ${next.reason}` : ""}`,
+	);
+	if (uokDiagnostics) {
+		lines.push(
+			`UOK: ${uokDiagnostics.verdict ?? "unknown"} (${uokDiagnostics.classification ?? "unknown"})`,
+		);
+	}
+	if (runtime.units.length > 0) {
+		lines.push("");
+		lines.push("Runtime units:");
+		for (const unit of runtime.units.slice(0, 8)) {
+			lines.push(`  ${unit.unitType} ${unit.unitId}: ${unit.status}`);
+		}
+	}
+	if (schedule) {
+		lines.push("");
+		lines.push(
+			`Schedule: ${schedule.pending_count} pending, ${schedule.overdue_count} overdue`,
+		);
+	}
+	if (state.registry.length > 0) {
+		lines.push("");
+		lines.push("Milestones:");
+		for (const milestone of state.registry) {
+			lines.push(`  ${milestone.id}: ${milestone.title} (${milestone.status})`);
+		}
+	}
+	return lines.join("\n");
+}
+
+/**
+ * Handle `sf headless status` without spawning the interactive RPC child.
+ *
+ * Purpose: avoid the long-standing v2 init timeout for a command whose answer
+ * is fully available from DB-backed project state.
+ *
+ * Consumer: runHeadlessOnce direct-command bypass.
+ */
+export async function handleHeadlessStatus(
+	basePath: string,
+	options: { json?: boolean } = {},
+): Promise<HeadlessStatusResult> {
+	const snapshot = await buildQuerySnapshot(basePath);
+	if (options.json) {
+		process.stdout.write(JSON.stringify(snapshot) + "\n");
+	} else {
+		process.stdout.write(formatHeadlessStatus(snapshot) + "\n");
+	}
+	return { exitCode: 0, data: snapshot };
+}
diff --git a/src/headless.ts b/src/headless.ts
index c13d4bcab..41cc91fcf 100644
--- a/src/headless.ts
+++ b/src/headless.ts
@@ -861,6 +861,22 @@ async function runHeadlessOnce(
 		return { exitCode: result.exitCode, interrupted: false, timedOut: false };
 	}
 
+	// Generic headless status: read-only project snapshot. This deliberately
+	// bypasses the interactive RPC/v2 path because `/status` opens a TUI overlay
+	// in interactive mode and can hang waiting for protocol init in headless.
+	if (options.command === "status") {
+		const { handleHeadlessStatus } = await import("./headless-status.js");
+		const wantsJson =
+			options.json ||
+			options.outputFormat === "json" ||
+			options.outputFormat === "stream-json" ||
+			options.commandArgs.includes("--json");
+		const result = await handleHeadlessStatus(process.cwd(), {
+			json: wantsJson,
+		});
+		return { exitCode: result.exitCode, interrupted: false, timedOut: false };
+	}
+
 	// Reflect: assemble the SF reflection corpus snapshot (open + recent
 	// self-feedback, recent commits, milestone state, validation files,
 	// prior report) and emit either the rendered prompt brief (default) or
diff --git a/src/resources/extensions/sf/memory-extractor.js b/src/resources/extensions/sf/memory-extractor.js
index 435ca4f05..0522008d9 100644
--- a/src/resources/extensions/sf/memory-extractor.js
+++ b/src/resources/extensions/sf/memory-extractor.js
@@ -149,19 +149,25 @@ function buildExtractionUserPrompt(
 /**
  * Extract assistant message text from activity JSONL.
  * Returns concatenated text content from assistant role entries.
+ *
+ * Truncation strategy (changed from front-preserving): the front of
+ * an execute-task transcript carries the goal/context (already in
+ * the prompt), while the tail carries verification, final summary,
+ * and the decisions actually committed — the highest-signal slices
+ * for memory extraction. Front-preserving truncation discarded
+ * exactly the content the LLM is best at extracting from.
+ *
+ * The fix: collect every assistant text block first, then if the
+ * total exceeds maxChars, keep ~25% from the front (goal echo,
+ * early decisions) and ~75% from the back (where the durable signal
+ * lives). The two halves are joined with a `[…truncated middle…]`
+ * marker so the LLM knows content was elided.
  */
 function extractTranscriptFromActivity(raw, maxChars = 30_000) {
 	const lines = raw.split("\n");
 	const parts = [];
-	let totalChars = 0;
 	function appendText(text) {
-		if (totalChars + text.length > maxChars) {
-			parts.push(text.substring(0, maxChars - totalChars));
-			return false; // signal stop
-		}
 		parts.push(text);
-		totalChars += text.length;
-		return true;
 	}
 	for (const line of lines) {
 		if (!line.trim()) continue;
@@ -178,21 +184,44 @@ function extractTranscriptFromActivity(raw, maxChars = 30_000) {
 			if (Array.isArray(entry.content)) {
 				for (const block of entry.content) {
 					if (block.type === "text" && block.text) {
-						if (!appendText(block.text)) return parts.join("\n\n");
+						appendText(block.text);
 					}
 				}
 			} else if (typeof entry.content === "string" && entry.content) {
-				if (!appendText(entry.content)) return parts.join("\n\n");
+				appendText(entry.content);
 			}
 			// Also read plain text/content field on custom_message entries
 			if (entry.text && typeof entry.text === "string") {
-				if (!appendText(entry.text)) return parts.join("\n\n");
+				appendText(entry.text);
 			}
 		} catch {
 			// Skip malformed lines
 		}
 	}
-	return parts.join("\n\n");
+	const joined = parts.join("\n\n");
+	if (joined.length <= maxChars) return joined;
+	return truncateMiddle(joined, maxChars);
+}
+
+/**
+ * Keep N from front + M from back of `text`, drop the middle.
+ *
+ * Split ratio is back-weighted (default 25/75) because in an
+ * execute-task transcript the durable signal — what was committed,
+ * what verified, what the operator decided — concentrates at the
+ * tail. The front share keeps enough goal/early-reasoning context
+ * to anchor the extraction; the back share keeps the outcomes.
+ *
+ * Exported (named export below) for the dedicated unit test that
+ * pins the ratio + the truncation marker.
+ */
+export function truncateMiddle(text, maxChars, frontShare = 0.25) {
+	if (text.length <= maxChars) return text;
+	const marker = "\n\n[…truncated middle…]\n\n";
+	const budget = Math.max(0, maxChars - marker.length);
+	const frontLen = Math.floor(budget * frontShare);
+	const backLen = budget - frontLen;
+	return text.slice(0, frontLen) + marker + text.slice(text.length - backLen);
 }
 // ─── Response Parsing ───────────────────────────────────────────────────────
 /**
diff --git a/src/resources/extensions/sf/memory-store.js b/src/resources/extensions/sf/memory-store.js
index 35400fcf5..33e22a8fb 100644
--- a/src/resources/extensions/sf/memory-store.js
+++ b/src/resources/extensions/sf/memory-store.js
@@ -5,6 +5,7 @@
 import { createMemoryRelation } from "./memory-relations.js";
 import {
 	_getAdapter,
+	computeStaticMemoryScore,
 	decayMemoriesBefore,
 	deleteMemoryEmbedding,
 	incrementMemoryHitCount,
@@ -78,7 +79,10 @@ function rankMemoriesByLexicalQuery(memories, query, limit) {
 				0,
 			);
 			const lexicalScore = lexicalHits / queryTokens.length;
-			const staticScore = memory.confidence * (1 + memory.hit_count * 0.1);
+			const staticScore = computeStaticMemoryScore(
+				memory.confidence,
+				memory.hit_count,
+			);
 			return {
 				memory,
 				index,
@@ -125,8 +129,13 @@ export function getActiveMemories() {
 	}
 }
 /**
- * Get active memories ordered by ranking score: confidence * (1 + hit_count * 0.1).
+ * Get active memories ordered by ranking score (computeStaticMemoryScore).
  * Higher-scored memories are more relevant and frequently confirmed.
+ *
+ * Sorting happens in JS because the canonical scoring formula uses
+ * log(), which the SQLite adapter doesn't reliably ship as a function
+ * across builds. The pool is bounded by `WHERE superseded_by IS NULL`
+ * which for typical projects (10s–1000s of rows) is cheap to sort.
  */
 export function getActiveMemoriesRanked(limit = 30) {
 	if (!isDbAvailable()) return [];
@@ -134,12 +143,17 @@ export function getActiveMemoriesRanked(limit = 30) {
 	if (!adapter) return [];
 	try {
 		const rows = adapter
-			.prepare(`SELECT * FROM memories
-       WHERE superseded_by IS NULL
-       ORDER BY (confidence * (1.0 + hit_count * 0.1)) DESC
-       LIMIT :limit`)
-			.all({ ":limit": limit });
-		return rows.map(rowToMemory);
+			.prepare(`SELECT * FROM memories WHERE superseded_by IS NULL`)
+			.all();
+		return rows
+			.map(rowToMemory)
+			.map((m) => ({
+				m,
+				score: computeStaticMemoryScore(m.confidence, m.hit_count),
+			}))
+			.sort((a, b) => b.score - a.score)
+			.slice(0, limit)
+			.map((entry) => entry.m);
 	} catch {
 		return [];
 	}
@@ -211,7 +225,7 @@ export async function getRelevantMemoriesRanked(query, limit = 10) {
 		let ranked = rankMemoriesByEmbedding(
 			mergedPool.map((m) => ({
 				id: m.id,
-				staticScore: m.confidence * (1 + m.hit_count * 0.1),
+				staticScore: computeStaticMemoryScore(m.confidence, m.hit_count),
 			})),
 			queryVec,
 			embeddingMap,
diff --git a/src/resources/extensions/sf/sf-db/sf-db-memory.js b/src/resources/extensions/sf/sf-db/sf-db-memory.js
index ac3cdbaa9..7b25aced9 100644
--- a/src/resources/extensions/sf/sf-db/sf-db-memory.js
+++ b/src/resources/extensions/sf/sf-db/sf-db-memory.js
@@ -1,6 +1,33 @@
 import { SF_STALE_STATE, SFError } from "../errors.js";
 import { _getAdapter, intBool, parseJsonObject } from "./sf-db-core.js";
 
+/**
+ * Canonical static-score formula for an active memory.
+ *
+ * Previously: confidence * (1 + hit_count * 0.1). That's linear in
+ * hit_count, which compounds: a memory that surfaces once gets
+ * hit_count++, ranks higher, surfaces again, ranks higher still.
+ * Self-reinforcing popularity, not relevance.
+ *
+ * Now: confidence * (1 + log(1 + hit_count) * 0.5). Same value at
+ * hit_count=0 (factor 1.0), close at low counts (hit_count=2 ≈ 1.55
+ * vs old 1.2), flattens fast (hit_count=10 ≈ 2.20 vs old 2.0), and
+ * stays bounded under runaway hits (hit_count=100 ≈ 3.30 vs old 11.0).
+ *
+ * Used by: memory-store ranked queries, memory-tools search ranker,
+ * supersedeLowestRankedMemories pruner (this file). One formula
+ * everywhere — no SQL/JS skew possible.
+ *
+ * Confidence in [0..1] by convention; hit_count non-negative integer.
+ * Defensive against NaN/negative inputs because the DB column is
+ * free-form numeric.
+ */
+export function computeStaticMemoryScore(confidence, hitCount) {
+	const c = Number.isFinite(confidence) ? Math.max(0, confidence) : 0;
+	const h = Number.isFinite(hitCount) ? Math.max(0, hitCount) : 0;
+	return c * (1 + Math.log(1 + h) * 0.5);
+}
+
 export function getActiveMemories({ category, limit = 200 } = {}) {
 	const currentDb = _getAdapter();
 	if (!currentDb) return [];
@@ -198,15 +225,38 @@ export function expireStaleMemories(unstartedTtlDays = 28, maxTtlDays = 90) {
 export function supersedeLowestRankedMemories(limit, now) {
 	const currentDb = _getAdapter();
 	if (!currentDb) throw new SFError(SF_STALE_STATE, "sf-db: No database open");
+	if (!Number.isFinite(limit) || limit <= 0) return;
+	// JS-side ranking via computeStaticMemoryScore so the formula is
+	// shared with memory-store's read paths. The candidate pool is
+	// bounded by the same SUPERSEDE LIMIT — fetch just enough to know
+	// which ids are the lowest-scoring.
+	const candidates = currentDb
+		.prepare(
+			`SELECT id, confidence, hit_count FROM memories WHERE superseded_by IS NULL`,
+		)
+		.all();
+	if (candidates.length === 0) return;
+	const ranked = candidates
+		.map((r) => ({
+			id: r.id,
+			score: computeStaticMemoryScore(r.confidence, r.hit_count),
+		}))
+		.sort((a, b) => a.score - b.score)
+		.slice(0, limit);
+	if (ranked.length === 0) return;
+	// IN-list with parameterised placeholders, one per id, so we never
+	// build the SQL with string concat.
+	const placeholders = ranked.map((_, i) => `:id${i}`).join(", ");
+	const params = { ":now": now };
+	ranked.forEach((r, i) => {
+		params[`:id${i}`] = r.id;
+	});
 	currentDb
-		.prepare(`UPDATE memories SET superseded_by = 'CAP_EXCEEDED', updated_at = :now
-     WHERE id IN (
-       SELECT id FROM memories
-       WHERE superseded_by IS NULL
-       ORDER BY (confidence * (1.0 + hit_count * 0.1)) ASC
-       LIMIT :limit
-     )`)
-		.run({ ":now": now, ":limit": limit });
+		.prepare(
+			`UPDATE memories SET superseded_by = 'CAP_EXCEEDED', updated_at = :now
+       WHERE id IN (${placeholders})`,
+		)
+		.run(params);
 }
 
 export function insertMemorySourceRow(args) {
diff --git a/src/resources/extensions/sf/tests/memory-static-score.test.mjs b/src/resources/extensions/sf/tests/memory-static-score.test.mjs
new file mode 100644
index 000000000..bcccb2581
--- /dev/null
+++ b/src/resources/extensions/sf/tests/memory-static-score.test.mjs
@@ -0,0 +1,79 @@
+/**
+ * Test the canonical static memory score formula.
+ *
+ * Previous formula `confidence * (1 + hit_count * 0.1)` was linear
+ * in hit_count → self-reinforcing popularity (a memory that surfaces
+ * once gets hit_count++, ranks higher, surfaces again, hits again...).
+ *
+ * Replacement `confidence * (1 + log(1 + hit_count) * 0.5)` keeps
+ * the curve at hit_count=0 identical, climbs noticeably at low hits
+ * (1.55× at h=2 vs 1.2× before — old was *too* flat there), and
+ * flattens fast (3.30× at h=100 vs the runaway 11.0× under the
+ * linear formula).
+ *
+ * This test pins:
+ *   - identity at h=0
+ *   - monotonic non-decreasing in h
+ *   - bounded growth (no value above an upper guardrail)
+ *   - defensiveness against NaN / negative inputs
+ */
+
+import assert from "node:assert/strict";
+import { test } from "vitest";
+import { computeStaticMemoryScore } from "../sf-db/sf-db-memory.js";
+
+test("identity at hit_count=0", () => {
+	assert.equal(computeStaticMemoryScore(0.8, 0), 0.8);
+	assert.equal(computeStaticMemoryScore(0.5, 0), 0.5);
+	assert.equal(computeStaticMemoryScore(1.0, 0), 1.0);
+});
+
+test("monotonic non-decreasing in hit_count", () => {
+	const c = 0.7;
+	let prev = computeStaticMemoryScore(c, 0);
+	for (let h = 1; h <= 1000; h++) {
+		const curr = computeStaticMemoryScore(c, h);
+		assert.ok(curr >= prev, `non-monotonic at h=${h}: ${curr} < ${prev}`);
+		prev = curr;
+	}
+});
+
+test("growth is bounded under runaway hit counts (log-shaped)", () => {
+	// Hand-computed pins: confidence * (1 + log(1+h)*0.5).
+	// h=100 → 1 + log(101)*0.5 ≈ 1 + 2.3081 ≈ 3.308
+	// h=10000 → 1 + log(10001)*0.5 ≈ 1 + 4.605 ≈ 5.605
+	const heavy = computeStaticMemoryScore(1.0, 10000);
+	assert.ok(heavy < 7, `score at h=10000 too high: ${heavy}`);
+	const veryHeavy = computeStaticMemoryScore(1.0, 1_000_000);
+	assert.ok(veryHeavy < 10, `score at h=1M too high: ${veryHeavy}`);
+});
+
+test("monotonic in confidence at fixed hit_count", () => {
+	for (let h = 0; h <= 100; h += 10) {
+		const low = computeStaticMemoryScore(0.2, h);
+		const mid = computeStaticMemoryScore(0.5, h);
+		const high = computeStaticMemoryScore(0.9, h);
+		assert.ok(low <= mid, `low<=mid at h=${h}: ${low} ${mid}`);
+		assert.ok(mid <= high, `mid<=high at h=${h}: ${mid} ${high}`);
+	}
+});
+
+test("NaN / negative inputs collapse to 0 rather than poisoning sorts", () => {
+	assert.equal(computeStaticMemoryScore(Number.NaN, 5), 0);
+	assert.equal(computeStaticMemoryScore(-0.5, 5), 0);
+	// Negative hit_count clamps to 0 → returns base confidence.
+	assert.equal(computeStaticMemoryScore(0.7, -3), 0.7);
+	assert.equal(computeStaticMemoryScore(0.7, Number.NaN), 0.7);
+});
+
+test("regression: new formula tames the runaway compared to linear", () => {
+	// Old: confidence * (1 + hit_count * 0.1) → at h=100 was 11×.
+	// New: cap at 1M hits stays under 10× (effective ceiling for any
+	// realistic memory). This pins the central property of the fix.
+	const linearAtHundred = 0.5 * (1 + 100 * 0.1); // 5.5
+	const newAtHundred = computeStaticMemoryScore(0.5, 100);
+	assert.ok(
+		newAtHundred < linearAtHundred,
+		`new (${newAtHundred}) should be < linear (${linearAtHundred}) at h=100`,
+	);
+});
diff --git a/src/resources/extensions/sf/tests/memory-truncate-middle.test.mjs b/src/resources/extensions/sf/tests/memory-truncate-middle.test.mjs
new file mode 100644
index 000000000..be647db71
--- /dev/null
+++ b/src/resources/extensions/sf/tests/memory-truncate-middle.test.mjs
@@ -0,0 +1,80 @@
+/**
+ * Test the front+back transcript truncation in memory-extractor.
+ *
+ * Previously: front-preserving truncation dropped the tail, where
+ * the highest-signal content for memory extraction lives (final
+ * summary, what got committed, verification outcome).
+ *
+ * Now: keep ~25% from front + ~75% from back with a marker in
+ * between. This test pins the invariants:
+ *   - identity for short inputs
+ *   - both ends survive when text exceeds cap
+ *   - the marker is present so the LLM knows about the elision
+ *   - output length never exceeds maxChars
+ */
+
+import assert from "node:assert/strict";
+import { test } from "vitest";
+import { truncateMiddle } from "../memory-extractor.js";
+
+test("identity when text fits", () => {
+	assert.equal(truncateMiddle("hello world", 30_000), "hello world");
+	assert.equal(truncateMiddle("", 30_000), "");
+});
+
+test("keeps front and back, drops middle, inserts marker", () => {
+	// Pick sizes so the surviving front + back ≪ middle, so the
+	// middle is meaningfully dropped (not just partly nibbled).
+	const front = "FRONT".repeat(1000); // 5K
+	const middle = "MIDDLE".repeat(20_000); // 120K
+	const back = "BACK".repeat(1000); // 4K
+	const text = front + middle + back; // 129K
+	const out = truncateMiddle(text, 10_000);
+	assert.ok(out.length <= 10_000, `output too long: ${out.length}`);
+	assert.match(out, /\[…truncated middle…\]/, "marker missing");
+	// Original front survives (first slice of FRONT pattern).
+	assert.ok(out.startsWith("FRONT"), "front not preserved");
+	// Original back survives (last slice ends with BACK pattern).
+	assert.ok(out.endsWith("BACK"), "back not preserved");
+	// Most of the giant middle is dropped — output is dramatically
+	// smaller than the input (>10× compression).
+	assert.ok(out.length < text.length / 10, `expected heavy compression`);
+});
+
+test("back-weighted split (75% back vs 25% front by default)", () => {
+	// Construct a text where each char identifies its position bucket.
+	const text = "F".repeat(10_000) + "M".repeat(10_000) + "B".repeat(10_000); // 30K
+	const out = truncateMiddle(text, 10_000);
+	const [head, tail] = out.split("[…truncated middle…]");
+	// Front share ≈ 25% of (10_000 - marker_len) ≈ ~2.5K. Back ≈ 7.5K.
+	assert.ok(
+		head.length < tail.length,
+		`expected back to be larger than front: front=${head.length} back=${tail.length}`,
+	);
+	assert.ok(
+		tail.length > head.length * 2,
+		`expected back to be > 2× front: front=${head.length} back=${tail.length}`,
+	);
+});
+
+test("respects custom frontShare", () => {
+	const text = "X".repeat(20_000);
+	const out = truncateMiddle(text, 5_000, 0.5);
+	const [head, tail] = out.split("[…truncated middle…]");
+	// 50/50 split.
+	assert.ok(
+		Math.abs(head.length - tail.length) <= 1,
+		`50/50 split should be balanced: front=${head.length} back=${tail.length}`,
+	);
+});
+
+test("output length never exceeds maxChars even with marker", () => {
+	for (const cap of [100, 500, 1_000, 30_000]) {
+		const text = "Z".repeat(cap * 3);
+		const out = truncateMiddle(text, cap);
+		assert.ok(
+			out.length <= cap,
+			`cap=${cap}: output length ${out.length} exceeded cap`,
+		);
+	}
+});
diff --git a/src/resources/extensions/sf/tools/memory-tools.js b/src/resources/extensions/sf/tools/memory-tools.js
index 51920240d..2119bdcd1 100644
--- a/src/resources/extensions/sf/tools/memory-tools.js
+++ b/src/resources/extensions/sf/tools/memory-tools.js
@@ -9,7 +9,7 @@ import {
 	getActiveMemoriesRanked,
 	reinforceMemory,
 } from "../memory-store.js";
-import { isDbAvailable } from "../sf-db.js";
+import { computeStaticMemoryScore, isDbAvailable } from "../sf-db.js";
 
 function dbUnavailable(operation) {
 	return {
@@ -97,7 +97,7 @@ export function executeMemoryQuery(params) {
 		});
 		const ranked = filtered.slice(0, k).map((memory) => ({
 			memory,
-			score: memory.confidence * (1 + memory.hit_count * 0.1),
+			score: computeStaticMemoryScore(memory.confidence, memory.hit_count),
 		}));
 		const hits = ranked.map((r) => ({
 			id: r.memory.id,
diff --git a/src/tests/headless-status.test.ts b/src/tests/headless-status.test.ts
new file mode 100644
index 000000000..5bcad8b03
--- /dev/null
+++ b/src/tests/headless-status.test.ts
@@ -0,0 +1,59 @@
+/**
+ * headless-status.test.ts — direct headless status rendering.
+ *
+ * Purpose: prevent `sf headless status` from regressing back to the
+ * interactive RPC/v2 path when it only needs DB-backed project state.
+ */
+import assert from "node:assert/strict";
+import { test } from "vitest";
+import type { QuerySnapshot } from "../headless-query.js";
+import { formatHeadlessStatus } from "../headless-status.js";
+
+function snapshot(overrides: Partial<QuerySnapshot> = {}): QuerySnapshot {
+	return {
+		schemaVersion: 1,
+		state: {
+			activeMilestone: { id: "M001", title: "Runtime Hardening" },
+			activeSlice: { id: "S01", title: "Headless Status" },
+			activeTask: { id: "T01", title: "Bypass v2 init" },
+			phase: "executing",
+			recentDecisions: [],
+			blockers: [],
+			nextAction: "Execute T01.",
+			registry: [{ id: "M001", title: "Runtime Hardening", status: "active" }],
+			requirements: {
+				active: 0,
+				validated: 0,
+				deferred: 0,
+				outOfScope: 0,
+				blocked: 0,
+				total: 0,
+			},
+			progress: {
+				milestones: { done: 0, total: 1 },
+				slices: { done: 0, total: 1 },
+				tasks: { done: 0, total: 1 },
+			},
+		},
+		next: {
+			action: "dispatch",
+			unitType: "execute-task",
+			unitId: "M001/S01/T01",
+		},
+		cost: { workers: [], total: 0 },
+		runtime: { units: [] },
+		uokDiagnostics: { verdict: "clear", classification: "healthy" },
+		schedule: { pending_count: 0, overdue_count: 0, due: [], upcoming: [] },
+		...overrides,
+	};
+}
+
+test("formatHeadlessStatus_when_snapshot_available_renders_text_without_rpc", () => {
+	const rendered = formatHeadlessStatus(snapshot());
+
+	assert.match(rendered, /^SF Status/);
+	assert.match(rendered, /Phase: executing/);
+	assert.match(rendered, /Active milestone: M001 - Runtime Hardening/);
+	assert.match(rendered, /Dispatch: dispatch execute-task M001\/S01\/T01/);
+	assert.match(rendered, /UOK: clear \(healthy\)/);
+});