perf(llm): load the two stream-guard modules in parallel

Both output-cap sites awaited liveDeadlines and streamFaultInjection in
sequence. The two imports are independent, so that costs an extra microtask
hop on every streamed turn for nothing (react-doctor
server-sequential-independent-await). This is the hot path: every answer
passes through it.

Both pairs are now a single Promise.all. The lone streamFaultInjection import
in trackCommit has no partner and is left alone.

Ported from the unmerged fix/stream-fault-injection (PR #464, closed). Its
sibling commit — the fault-injection switches themselves — is already on main
by another route; only this follow-up was left behind. A straight cherry-pick
would have conflicted: the streamChat site now also destructures
MAX_SUMMARY_OUTPUT_CHARS, which was added after that branch was cut, so the
first hunk had to be re-derived rather than replayed.

Behaviour is unchanged — same modules, same bindings, same order of use.

Validation:
  - Build validated on macOS: tsc -p electron/tsconfig.json --noEmit, 0 errors
  - Covered by automated tests: StreamFaultInjection 8/8 (including "both
    output-cap sites honour the test ceiling", which pins exactly these two
    sites), RunawayStreamOutputCap 20/20, CodingRegenCeiling 3/3,
    PostCommitNoProviderSwitch 10/10
  - Caveat: RunawayStreamOutputCap imports the built bundle from
    dist-electron, which predates this edit, so its runtime assertions
    exercised the pre-change build. Its source-assertion half did read the
    patched .ts. A rebuild would close that gap; the change is a semantically
    identical refactor, so the residual risk is low.
  - Platform-neutral: no platform-specific code path is touched.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JJS2xjF8uDPMa4EpXey5FL

Author: Hardening v2.7

electron/LLMHelper.ts

@@ -5529,8 +5529,17 @@ let isMultimodal = !!(imagePaths?.length);
     // The ceiling is per-SURFACE: the live cap is calibrated from what_to_answer
     // answers and is far too tight for a whole-meeting summary, which the batch
     // callers reach through streamChatLongForm.
-    const { MAX_STREAM_OUTPUT_CHARS, MAX_SUMMARY_OUTPUT_CHARS } = await import('./llm/liveDeadlines');
-    const { testOutputCharCeiling } = await import('./llm/streamFaultInjection');
+    // Loaded TOGETHER: the two imports are independent, so awaiting them in
+    // sequence costs an extra microtask hop on every streamed turn for nothing
+    // (react-doctor server-sequential-independent-await). This is the hot path —
+    // every answer passes through it.
+    const [
+      { MAX_STREAM_OUTPUT_CHARS, MAX_SUMMARY_OUTPUT_CHARS },
+      { testOutputCharCeiling },
+    ] = await Promise.all([
+      import('./llm/liveDeadlines'),
+      import('./llm/streamFaultInjection'),
+    ]);
     // Test switch (dev-only, opt-in) lets the cap be provoked without waiting
     // for a model to actually loop. Null in any packaged build.
     const outputCeiling = testOutputCharCeiling()
@@ -5629,8 +5638,12 @@ let isMultimodal = !!(imagePaths?.length);
     state: { chars: number; truncated?: boolean },
     label: string,
   ): AsyncGenerator<string, void, unknown> {
-    const { MAX_STREAM_OUTPUT_CHARS } = await import('./llm/liveDeadlines');
-    const { testOutputCharCeiling } = await import('./llm/streamFaultInjection');
+    // Same independent-await pair as streamChat above — one Promise.all rather
+    // than two sequential awaits on the per-chunk cap path.
+    const [{ MAX_STREAM_OUTPUT_CHARS }, { testOutputCharCeiling }] = await Promise.all([
+      import('./llm/liveDeadlines'),
+      import('./llm/streamFaultInjection'),
+    ]);
     const ceiling = testOutputCharCeiling() ?? MAX_STREAM_OUTPUT_CHARS;
     for await (const chunk of inner) {
       yield chunk;