Handle empty AI quote recommendations

Treat the "AI finished but recommended zero products" case as its own terminal SSE event instead of crashing through quote-creation as a failure.

1. Why this exists

The AI quoting pipeline emits SSE events as it walks an opportunity through stages (find similar accounts → recommend products → map SKUs → build quote). Previously there were two terminal events: complete (a real quote was built) and error (something blew up).

An empty recommendation — the AI ran clean but decided it didn't have enough product evidence for this account — fit neither bucket. The route would keep going into SpecService / quote creation with an empty products array, and downstream filtering would surface this as a quote-generation failure in the UI. Reviewers got a red error for what is actually a normal, low-confidence outcome.

2. Before / after pipeline shape

3. What changes, file by file

Server: new terminal event type

apps/server/src/dealops2/aiQuoting/pipeline/streamTypes.ts adds 'no_quote' to the AIQuoteEventType union and a new NoQuoteEvent shape carrying everything the UI needs to render without a follow-up fetch:

export interface NoQuoteEvent {
  type: 'no_quote';
  success: true;                     // not an error
  reason: string;                    // AI's overallReasoning prose
  confidence: 'high' | 'medium' | 'low';
  summary: {
    targetAccount: string;
    similarAccountsFound: number;
    aiProductsRecommended: number;
    mappedProducts: number;
  };
  timestamp: number;
}

The event is added to the AIQuoteEvent discriminated union so both ends stay type-checked.

Server: route short-circuit

apps/server/src/routes/aiQuoting.ts adds a guard immediately after the AI stage finishes and before SpecService spins up:

if (aiResult.payload.products.length === 0) {
  emitter.emit(
    createEvent<NoQuoteEvent>({
      type: 'no_quote',
      success: true,
      reason:
        aiResult.aiRecommendation?.overallReasoning ||
        'The AI did not find enough product evidence …',
      confidence: aiResult.aiRecommendation?.overallConfidence ?? 'low',
      summary: { /* targetAccount, similarAccountsFound, … */ },
    }),
  );
  res.end();
  return;
}

Note this runs after auditState.stage5 = aiResult.stage5, so the audit record still captures the AI's full output — only the pricing-spec / quote-build path is skipped.

Client hook: new status + state field

useAIQuoteStream.ts extends the state machine:

status:
  | 'idle'
  | 'connecting'
  | 'running'
  | 'completed'
  | 'no_quote'   // new
  | 'error'

noQuote: AIQuoteNoQuoteResult | null

isNoQuote: state.status === 'no_quote'

The SSE message handler gets a new case 'no_quote' branch that stores the payload and flips status. All three reset paths (initial, start, reset) zero out the new noQuote field alongside result.

Client modal: banner + parser + close-only footer

AIQuoteModal.tsx threads isNoQuote through HeaderBlock (subtitle becomes "No quote created") and renders a new <NoQuoteBanner> component. The footer condition expands from isError to isError || isNoQuote, giving the no-quote state its own close bar.

The interesting piece is the reasoning parser. The AI's overallReasoning tends to arrive as a flat blob with markdown-ish headings; parseNoQuoteReason splits it on Evidence: / Benchmark: / Strategy: headings, strips **bold** and bullet markers, and falls back to a single "Reason" section if no headings are found. Each section caps display at 3 lines.

4. What the no-quote banner looks like

An approximation of the rendered banner (real version is amber Tailwind classes on the client). Confidence pill is right-aligned, parsed reasoning sits below a divider.

5. How the parser handles messy AI prose

The parsing logic is defensive about the shape of the AI string:

1. Normalize \r\n → \n and trim.
2. Run a global regex for headings: /(?:^|\n)\s*\*{0,2}(Evidence|Benchmark|Strategy)\*{0,2}:?\s*/gi.
3. No matches? Wrap the whole string as a single "Reason" section.
4. Otherwise slice between consecutive heading matches, clean each line (strip **, leading -/*/•), drop empties.
5. Drop sections that ended up with zero lines.

Display tops out at 3 lines per section — if the AI rambles, the banner stays compact.

6. What it doesn't change

7. Risks and open questions