perf(newos4): delegated agent runs no longer serialise; a waiter can't steal a lock

Closes the second half of STEER-20261102-201045-ab08.

- chat() gated the main-turn lock on nothing, so orchestrate-parallel's fan-out
  (each sub-task calling chat() on the same instance with its own agentTag) ran
  strictly one at a time. Measured 2x300ms delegated runs: 617ms before, 306ms
  now; the two 200ms user turns still take 400ms (main path stays serialised).
- the 10s lock timeout called this._chatLock = false and printed "force
  released", which cleared a lock the waiter did not own - that released the
  main turn's lock while it was still running. A waiter now throws and the owner
  is the only writer.
- delegated runs no longer clear the user's cancel flag on entry (an agent start
  used to reset this._cancelled, disarming a Stop the user had pressed).
- tool-log entries carry the run they belong to and _recentToolLog(tag, n) is
  what feeds the run log line and _postRunReview, so concurrent runs no longer
  consume each other's entries.

Known limit, unchanged: _currentTag is a single field, so under true concurrency
an entry can be attributed to the wrong run (telemetry only - the answers are
per-session). Namespacing it properly needs per-call state threading.

Test: modules/test/test-chat-concurrency.js, 7 assertions, RED before
(H1 617ms, H3 force-released, H5 no tags) and GREEN now; maintain 24/24.

3 files changed, 63 insertions, 21 deletions

diff --git a/modules/core/core-bundle.js b/modules/core/core-bundle.js
index ef2d425..cf805a9 100644
--- a/modules/core/core-bundle.js
+++ b/modules/core/core-bundle.js
@@ -9699,18 +9699,24 @@ max_chars: { type: 'integer', description: 'Max characters of page content to re
                 if (!userMessage) {
                     this.terminal.print('Usage: ai <your-question>', 'warning'); return;
                 }
-                this._cancelled = false;
-                this.terminal.setAiBusy(true);
+                // A delegated run owns its own session (see _isAgentRun below) and must not queue
+                // behind the main turn: orchestrate-parallel's fan-out used to run strictly one
+                // sub-task at a time, and a waiter that hit the timeout cleared a lock it did not
+                // own — which released the main turn's lock while it was still running.
+                const _agentTag = (opts && opts.agentTag) || null;
+                if (!_agentTag) this._cancelled = false;
                 const lockStart = Date.now();
-                while (this._chatLock) {
+                while (!_agentTag && this._chatLock) {
                     if (Date.now() - lockStart > this._chatLockTimeoutMs) {
-                        this._chatLock = false;
-                        this.terminal.print('⚠️ Chat lock timed out — force released', 'warning');
-                        break;
+                        throw new Error("chat lock busy after " + Math.round(this._chatLockTimeoutMs / 1000)
+                            + "s — refusing to release another turn's lock");
                     }
                     await new Promise(r => setTimeout(r, 50));
                 }
-                this._chatLock = true;
+                if (!_agentTag) this._chatLock = true;
+                const _prevTag = this._currentTag;
+                this._currentTag = _agentTag || 'main';
+                this.terminal.setAiBusy(true);
                 const taskStart = Date.now();   // predictor durations exclude chat-lock wait
                 this.terminal.markDirty();
                 // Auto-route context based on intent
@@ -9867,7 +9873,7 @@ max_chars: { type: 'integer', description: 'Max characters of page content to re
                     try {
                         var om = this.terminal ? this.terminal.ollamaManager : null;
                         var usageData = om && om._usageLog && om._usageLog.length > 0 ? om._usageLog[om._usageLog.length - 1] : null;
-                        this._postRunReview(loopResult, this._toolCallLog.slice(-(loopResult.toolTurns || 0)), usageData);
+                        this._postRunReview(loopResult, this._recentToolLog(_agentTag || 'main', loopResult.toolTurns || 0), usageData);
                     } catch (e) { /* self-improvement is best-effort */ }
                 }
                 // Auto-create skills from repeated errors
@@ -9884,7 +9890,8 @@ max_chars: { type: 'integer', description: 'Max characters of page content to re
                     setTimeout(() => this.terminal.playNotificationSound(), 100);
                 }
                 } finally {
-                    this._chatLock = false;
+                    this._currentTag = _prevTag;
+                    if (!_agentTag) this._chatLock = false;
                     this.terminal.setAiBusy(false);
                     // Persist only the user's conversation — agent sessions are ephemeral.
                     if (!_isAgentRun) this._saveContext();
@@ -9894,7 +9901,7 @@ max_chars: { type: 'integer', description: 'Max characters of page content to re
                         // `recent` was referenced but never declared: the bare identifier threw
                         // ReferenceError inside this finally block, rejecting chat() on every
                         // tool-using turn (shipped in 32608bc, caught by the caller sweep).
-                        const recent = this._toolCallLog.slice(-(loopResult.toolTurns || 0));
+                        const recent = this._recentToolLog(_agentTag || 'main', loopResult.toolTurns || 0);
                         const toolNames = recent.map(function(e) { return e.name; }).join(', ');
                         this.terminal.log('INFO', 'AI', 'Used ' + loopResult.toolTurns + ' tool(s): ' + toolNames
                             + ' · returned ' + Math.round(recent.reduce((a, e) => a + e.returnedBytes, 0) / 1024) + 'KB'
@@ -10231,6 +10238,19 @@ max_chars: { type: 'integer', description: 'Max characters of page content to re
             // went 6 → 112 705 tokens across 8 model calls (maxContextTokens is 12 000). Two
             // limits keep a single turn inside the window: clip each observation as it arrives,
             // and elide the oldest ones once the running total passes the budget.
+            // A run may only see the tool calls it made: with concurrent delegated runs the
+            // raw ring buffer interleaves, and slicing it by toolTurns handed one agent
+            // another agent's entries.
+            _recentToolLog(tag, n) {
+                const want = tag || 'main';
+                const out = [];
+                const cap = Math.max(0, Number(n) || 0);
+                for (let i = this._toolCallLog.length - 1; i >= 0 && out.length < cap; i--) {
+                    const e = this._toolCallLog[i];
+                    if ((e.tag || 'main') === want) out.unshift(e);
+                }
+                return out;
+            }
             _recordToolCall(entry) {
                 const r = String(entry.result || '');
                 let truncated = false, contentChars = null;
@@ -10250,6 +10270,7 @@ max_chars: { type: 'integer', description: 'Max characters of page content to re
                     clippedBytes: Math.max(0, r.length - Number(entry.returnedBytes || r.length)),
                     truncated: truncated, durationMs: entry.durationMs || 0,
                     ok: !entry.error, errorKind: entry.errorKind || null,
+                    tag: entry.tag || this._currentTag || 'main',
                 };
                 this._toolCallLog.push(rec);
                 if (this._toolCallLog.length > 50) this._toolCallLog.shift();
diff --git a/modules/core/module-aiassist.js b/modules/core/module-aiassist.js
index 417e1b6..db4e691 100644
--- a/modules/core/module-aiassist.js
+++ b/modules/core/module-aiassist.js
@@ -824,18 +824,24 @@ max_chars: { type: 'integer', description: 'Max characters of page content to re
                 if (!userMessage) {
                     this.terminal.print('Usage: ai <your-question>', 'warning'); return;
                 }
-                this._cancelled = false;
-                this.terminal.setAiBusy(true);
+                // A delegated run owns its own session (see _isAgentRun below) and must not queue
+                // behind the main turn: orchestrate-parallel's fan-out used to run strictly one
+                // sub-task at a time, and a waiter that hit the timeout cleared a lock it did not
+                // own — which released the main turn's lock while it was still running.
+                const _agentTag = (opts && opts.agentTag) || null;
+                if (!_agentTag) this._cancelled = false;
                 const lockStart = Date.now();
-                while (this._chatLock) {
+                while (!_agentTag && this._chatLock) {
                     if (Date.now() - lockStart > this._chatLockTimeoutMs) {
-                        this._chatLock = false;
-                        this.terminal.print('⚠️ Chat lock timed out — force released', 'warning');
-                        break;
+                        throw new Error("chat lock busy after " + Math.round(this._chatLockTimeoutMs / 1000)
+                            + "s — refusing to release another turn's lock");
                     }
                     await new Promise(r => setTimeout(r, 50));
                 }
-                this._chatLock = true;
+                if (!_agentTag) this._chatLock = true;
+                const _prevTag = this._currentTag;
+                this._currentTag = _agentTag || 'main';
+                this.terminal.setAiBusy(true);
                 const taskStart = Date.now();   // predictor durations exclude chat-lock wait
                 this.terminal.markDirty();
                 // Auto-route context based on intent
@@ -992,7 +998,7 @@ max_chars: { type: 'integer', description: 'Max characters of page content to re
                     try {
                         var om = this.terminal ? this.terminal.ollamaManager : null;
                         var usageData = om && om._usageLog && om._usageLog.length > 0 ? om._usageLog[om._usageLog.length - 1] : null;
-                        this._postRunReview(loopResult, this._toolCallLog.slice(-(loopResult.toolTurns || 0)), usageData);
+                        this._postRunReview(loopResult, this._recentToolLog(_agentTag || 'main', loopResult.toolTurns || 0), usageData);
                     } catch (e) { /* self-improvement is best-effort */ }
                 }
                 // Auto-create skills from repeated errors
@@ -1009,7 +1015,8 @@ max_chars: { type: 'integer', description: 'Max characters of page content to re
                     setTimeout(() => this.terminal.playNotificationSound(), 100);
                 }
                 } finally {
-                    this._chatLock = false;
+                    this._currentTag = _prevTag;
+                    if (!_agentTag) this._chatLock = false;
                     this.terminal.setAiBusy(false);
                     // Persist only the user's conversation — agent sessions are ephemeral.
                     if (!_isAgentRun) this._saveContext();
@@ -1019,7 +1026,7 @@ max_chars: { type: 'integer', description: 'Max characters of page content to re
                         // `recent` was referenced but never declared: the bare identifier threw
                         // ReferenceError inside this finally block, rejecting chat() on every
                         // tool-using turn (shipped in 32608bc, caught by the caller sweep).
-                        const recent = this._toolCallLog.slice(-(loopResult.toolTurns || 0));
+                        const recent = this._recentToolLog(_agentTag || 'main', loopResult.toolTurns || 0);
                         const toolNames = recent.map(function(e) { return e.name; }).join(', ');
                         this.terminal.log('INFO', 'AI', 'Used ' + loopResult.toolTurns + ' tool(s): ' + toolNames
                             + ' · returned ' + Math.round(recent.reduce((a, e) => a + e.returnedBytes, 0) / 1024) + 'KB'
@@ -1356,6 +1363,19 @@ max_chars: { type: 'integer', description: 'Max characters of page content to re
             // went 6 → 112 705 tokens across 8 model calls (maxContextTokens is 12 000). Two
             // limits keep a single turn inside the window: clip each observation as it arrives,
             // and elide the oldest ones once the running total passes the budget.
+            // A run may only see the tool calls it made: with concurrent delegated runs the
+            // raw ring buffer interleaves, and slicing it by toolTurns handed one agent
+            // another agent's entries.
+            _recentToolLog(tag, n) {
+                const want = tag || 'main';
+                const out = [];
+                const cap = Math.max(0, Number(n) || 0);
+                for (let i = this._toolCallLog.length - 1; i >= 0 && out.length < cap; i--) {
+                    const e = this._toolCallLog[i];
+                    if ((e.tag || 'main') === want) out.unshift(e);
+                }
+                return out;
+            }
             _recordToolCall(entry) {
                 const r = String(entry.result || '');
                 let truncated = false, contentChars = null;
@@ -1375,6 +1395,7 @@ max_chars: { type: 'integer', description: 'Max characters of page content to re
                     clippedBytes: Math.max(0, r.length - Number(entry.returnedBytes || r.length)),
                     truncated: truncated, durationMs: entry.durationMs || 0,
                     ok: !entry.error, errorKind: entry.errorKind || null,
+                    tag: entry.tag || this._currentTag || 'main',
                 };
                 this._toolCallLog.push(rec);
                 if (this._toolCallLog.length > 50) this._toolCallLog.shift();
diff --git a/modules/core/module-loader.js b/modules/core/module-loader.js
index eb916d1..7206155 100644
--- a/modules/core/module-loader.js
+++ b/modules/core/module-loader.js
@@ -4,7 +4,7 @@
     // Core module bundle loaded synchronously in main.html.
     // This loader only handles version cache-busting, splash, terminal init, and optional ext modules.
 
-    const LATEST_VER = 'v4.6.138';
+    const LATEST_VER = 'v4.6.139';
     const CACHE = '?v=' + LATEST_VER;
     // Exposed so ext modules can register their own URL with the same cache-busting version
     // instead of an unversioned path the browser may already have cached (shell.js did).