diff --git a/CHANGELOG.md b/CHANGELOG.md
index 9fedd9d..10882fd 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -66,6 +66,8 @@ Cutting that release is tracked in
 
 ### Changed
 
+- Context-budget pruning now measures each candidate once, removes oversized
+  entries regardless of array position, and byte-budgets text fallback.
 - Removed stale screenshot and JSON conversion paths, honored per-operation
   Linux evaluation timeouts, and stopped advertising the compatibility-only
   `--json` flag.
diff --git a/apps/headless/Sources/HeadlessProtocol/AgentRuntime.swift b/apps/headless/Sources/HeadlessProtocol/AgentRuntime.swift
index 67199f5..f3c5560 100644
--- a/apps/headless/Sources/HeadlessProtocol/AgentRuntime.swift
+++ b/apps/headless/Sources/HeadlessProtocol/AgentRuntime.swift
@@ -371,31 +371,66 @@ if (!globalThis.__headlessAgent) {
           budget,
           budgetApplied: Boolean(budget)
         };
-        for (let pass = 0; pass < 2; pass += 1) {
-          const bytes = encodedBytes(result);
-          result.contextStats.encodedBytes = bytes;
-          result.contextStats.estimatedTokens = Math.ceil(bytes / 4);
+        const baseBytes = encodedBytes(result);
+        let bytes = baseBytes;
+        let estimatedTokens = Math.ceil(bytes / 4);
+        while (true) {
+          const measured = baseBytes + String(bytes).length - 1 + String(estimatedTokens).length - 1;
+          const measuredTokens = Math.ceil(measured / 4);
+          if (measured === bytes && measuredTokens === estimatedTokens) break;
+          bytes = measured;
+          estimatedTokens = measuredTokens;
         }
+        result.contextStats.encodedBytes = bytes;
+        result.contextStats.estimatedTokens = estimatedTokens;
+        return bytes;
       };
-      refresh();
-      if (budget) {
-        const maximumBytes = budget * 4;
-        while (encodedBytes(result) > maximumBytes) {
-          const populated = arrays.filter(key => result[key].length > 0);
-          if (populated.length === 0) {
-            if (typeof result.text === 'string' && result.text.length > 0) {
-              result.text = result.text.slice(0, Math.max(0, result.text.length - 256));
-              refresh();
-              continue;
-            }
-            break;
-          }
-          populated.sort((left, right) => encodedBytes(result[right][result[right].length - 1]) - encodedBytes(result[left][result[left].length - 1]));
-          result[populated[0]].pop();
-          refresh();
+      let currentBytes = refresh();
+      if (!budget) return result;
+
+      const maximumBytes = budget * 4;
+      if (currentBytes > maximumBytes) {
+        const candidates = arrays.flatMap(key => result[key].map((value, index) => ({
+          key, index, bytes: encodedBytes(value) + (result[key].length > 1 ? 1 : 0)
+        }))).sort((left, right) => right.bytes - left.bytes);
+        const removed = new Map(arrays.map(key => [key, new Set()]));
+        // Reserve a small allowance for omitted-count digit growth. Each item
+        // is encoded once; removal then preserves the order of retained items.
+        const targetBytes = Math.max(0, maximumBytes - 32);
+        for (const candidate of candidates) {
+          if (currentBytes <= targetBytes) break;
+          removed.get(candidate.key).add(candidate.index);
+          currentBytes -= candidate.bytes;
+        }
+        for (const key of arrays) {
+          const indexes = removed.get(key);
+          if (indexes.size > 0) result[key] = result[key].filter((_, index) => !indexes.has(index));
         }
+        currentBytes = refresh();
+      }
+
+      if (currentBytes > maximumBytes && typeof result.text === 'string' && result.text.length > 0) {
+        const textBytes = Math.max(0, encodedBytes(result.text) - 2);
+        const targetTextBytes = Math.max(0, textBytes - (currentBytes - maximumBytes) - 32);
+        let lower = 0;
+        let upper = result.text.length;
+        while (lower < upper) {
+          const middle = Math.ceil((lower + upper) / 2);
+          const prefixBytes = Math.max(0, encodedBytes(result.text.slice(0, middle)) - 2);
+          if (prefixBytes <= targetTextBytes) lower = middle;
+          else upper = middle - 1;
+        }
+        result.text = result.text.slice(0, lower);
+        currentBytes = refresh();
+      }
+
+      // Metadata is deliberately bounded, but keep the budget fail-closed if
+      // its final digit widths ever outgrow the allowance above.
+      if (currentBytes > maximumBytes) {
+        for (const key of arrays) result[key] = [];
+        if (typeof result.text === 'string') result.text = '';
+        refresh();
       }
-      refresh();
       return result;
     };
     const snapshot = (interactiveOnly, includeText, options = {}) => {
diff --git a/apps/headless/Tests/agent-runtime.test.mjs b/apps/headless/Tests/agent-runtime.test.mjs
index 2064416..af92ec8 100644
--- a/apps/headless/Tests/agent-runtime.test.mjs
+++ b/apps/headless/Tests/agent-runtime.test.mjs
@@ -99,6 +99,44 @@ const scopedActions = agent.snapshot(false, false, {
 assert(scopedActions.elements.some(element => element.name === 'Copy authentication command'));
 assert(scopedActions.elements.every(element => element.actions.length > 0));
 
+// Budget pruning measures every candidate once and removes the largest item
+// wherever it appears, without discarding smaller entries that follow it.
+const pruningFixture = window.document.createElement('section');
+pruningFixture.innerHTML = `
+  <p>Budget prefix should survive.</p>
+  <p>${'🚧'.repeat(300)}</p>
+  <p>Budget suffix should survive.</p>
+`;
+for (const [index, paragraph] of [...pruningFixture.querySelectorAll('p')].entries()) {
+  paragraph.getBoundingClientRect = () => ({
+    x: 24, y: 100 + index * 50, top: 100 + index * 50, left: 24,
+    right: 824, bottom: 148 + index * 50, width: 800, height: 48,
+  });
+}
+window.document.getElementById('content').prepend(pruningFixture);
+let encodeCalls = 0;
+const NativeTextEncoder = window.TextEncoder;
+window.TextEncoder = class CountingTextEncoder extends NativeTextEncoder {
+  encode(value) {
+    encodeCalls += 1;
+    return super.encode(value);
+  }
+};
+const middlePruned = agent.snapshot(false, false, {context: 'text', limit: 4, budget: 256});
+window.TextEncoder = NativeTextEncoder;
+assert(
+  middlePruned.snippets.some(item => item.text === 'Budget suffix should survive.'),
+  JSON.stringify(middlePruned),
+);
+assert(!middlePruned.snippets.some(item => item.text.includes('🚧')));
+assert(encodeCalls <= 8, 'budget pruning should encode the result and each candidate only once');
+assert(middlePruned.contextStats.encodedBytes <= 1024);
+assert.equal(
+  new TextEncoder().encode(JSON.stringify(middlePruned)).length,
+  middlePruned.contextStats.encodedBytes,
+  'reported bytes should include the finalized context statistics',
+);
+
 assert.throws(
   () => agent.snapshot(false, false, {context: 'text', within: '@r999999'}),
   error => error.headlessCode === 'REGION_NOT_FOUND' && /REGION_NOT_FOUND/.test(error.message),
@@ -256,6 +294,10 @@ const budgetedText = agent.snapshot(false, true, {context: 'full', limit: 1, bud
 assert(budgetedText.contextStats.encodedBytes <= 1024);
 assert(budgetedText.text.length < 30000, 'text fallback should be shortened to fit the budget');
 assert.equal(budgetedText.contextStats.budgetApplied, true);
+assert.equal(
+  new TextEncoder().encode(JSON.stringify(budgetedText)).length,
+  budgetedText.contextStats.encodedBytes,
+);
 
 console.log(JSON.stringify({
   selectedRegion: targetRegion.ref,
diff --git a/docs/roadmap/improvements-backlog.md b/docs/roadmap/improvements-backlog.md
index 12718d8..cd602bc 100644
--- a/docs/roadmap/improvements-backlog.md
+++ b/docs/roadmap/improvements-backlog.md
@@ -191,11 +191,14 @@ Foundation conversion paths, made both hosts share `JSONValue.foundationValue`,
 honored the bounded Linux evaluation timeout, deleted the unreachable PDF
 branch, and hid the backward-compatible no-op `--json` parser flag from help.
 
-**B5. `pruneToBudget` quality.** ([#25](https://github.com/LockInTime/headless/issues/25)) Hand-rolled 2-pass fixed point
+**B5. `pruneToBudget` quality.** ([#25](https://github.com/LockInTime/headless/issues/25)) ~~Hand-rolled 2-pass fixed point
 (`HP/AgentRuntime.swift:348-352`), O(n²) re-encoding per trim, pop-largest-
 *last*-element heuristic misses large mid-array items
 (`AgentRuntime.swift:367-369`), text-chop fallback untested. Rework with a
-size-estimating single pass; add unit tests in the jsdom suite.
+size-estimating single pass; add unit tests in the jsdom suite.~~ **Done:**
+each candidate is measured once, largest entries are pruned regardless of
+position while retained order is stable, text uses a byte-budgeted prefix
+search, and jsdom locks the mid-array and text-fallback cases.
 
 **B6. Declared capability matrix.** ([#26](https://github.com/LockInTime/headless/issues/26)) Silent per-platform divergences to either
 fix or promote to declared differences asserted in tests: PDF raster (macOS,