Tool audit sweep: 4 BLOCKERS, 12 HIGH, 6 MEDIUM fixed across 23 files

After 4 parallel agent audits covering 47 tool surfaces, found and fixed:

BLOCKERS (4):
- Display banner CharBadge used 30/90 hardcoded limits for every size
  regardless of which banner (Mobile Banner = 22/28, Wide Skyscraper = 25/80,
  etc.). Imports DISPLAY_SIZES and looks up per-size h_max/d_max now.
- Google Shopping policy_warnings render was unreachable: outer `?.length`
  truthy gate meant the zero-warnings "clean" message never fired. Now uses
  Array.isArray() check.
- Google Shopping `category` field smart-fill aliased to brain.industry,
  pre-populating "E-commerce" into the product-taxonomy field. Renamed to
  product_category so smart-fill leaves it blank for explicit user input.
- Hashtags tool: HashtagOutput used config.initial.platform (frozen as
  "instagram") to key into the recommended-set bucket. Non-IG platforms
  always showed empty recommendations. Now discovers the key dynamically
  with Object.keys().find(k => k.startsWith("recommended_set_for_")).

HIGH (12):
- Unconditional "IF AN IMAGE IS ATTACHED" prompt blocks in 5 optimizer
  prompts (ad-fatigue, bid-strategy, ctr-optimizer, keyword-strategy,
  landing-page) — now gated on `input.<screenshot_field>` matching the
  fix already applied to audience-targeting. Each Input interface gains
  its `?_screenshot?: unknown` field.
- LinkedIn `format.replace("_", " ")` only replaced first underscore;
  "lead_gen_form" became "lead gen_form". Now uses /_/g regex.
- YouTube copy-button dropped beat timestamps; in-stream script lines
  copied without their time markers. Now includes [t] in the copy text.
- Budget-planner buildTitle hardcoded `$${total_monthly}/mo`; INR user
  would see "$50000" in saved history. Dropped the `$`.
- Budget-planner renderer + budget renderer: 12 hardcoded `$` literals
  in AI-generated figures (waste, recovery, monthly_usd, etc.). All now
  use getCurrency().symbol. Prompts also prefix amounts with currency code
  so the AI generates figures in the user's chosen currency.
- Audience optimizer label "Budget / mo (USD)" → "Budget / mo" with
  generic "Your selected currency" hint.
- Suggestions auto-run race: setTimeout(() => run(), 400) captured stale
  null brain via closure, causing "No active Brand Brain" error on load
  when a brain existed. New runWithBrain(b) accepts the brain directly.
- Launch wizard fallback retry dropped master abort signal; Stop button
  couldn't cancel during fallback. Now uses anySignal(args.signal,
  fallbackCtl.signal).
- Wizard + GeneratorShell + suggestions + competitors + frameworks +
  learn/[concept] + batch.tsx: ALL now use `res.text || stream.text`
  fallback so empty res.text from Gemini doesn't produce blank saved ads.
- Batch.tsx: no AbortController, no Stop button — N parallel calls
  couldn't be cancelled. Now has master signal threaded to each llmStream
  + Stop button visible while running.
- learn/[concept]/Client.tsx: no AbortController; streaming continued
  writing to state after unmount. Added abortRef + useEffect cleanup.
- BrandBrainForm: extract() did direct spread merge, destroying manual
  edits. Now uses mergeFillEmpty. Also wired AbortController to all 3
  AI call paths. save() redirect changed from "/" to "/brand".
- Landing-page tool: URL field was a dead string the AI couldn't fetch.
  Updated label + hint to make this explicit; user pastes copy in the
  textarea below.
- Spark-ads + suggestions: `$` hardcoded in scale-rules cap and campaign
  budget display. Both now use getCurrency().symbol.

MEDIUM (6):
- campaign-kit prompt now imports BANNED_WORDS_RULE + HONESTY_CLAUSE
  matching meta-ads / google-ads / linkedin-ads.
- content-calendar prompt computes estimated entry count (days × platforms
  × cadence/wk) and instructs a HARD_ENTRY_CAP=35 with truncation_notice
  to prevent silent JSON truncation on 1-month multi-platform requests
  that previously exceeded the 6500-token budget.
- google-shopping price field placeholder no longer says "$24.99";
  prompt guard: "(not provided — do not invent pricing)".
- Display, hashtags, google-shopping minor cleanup.

Verification: typecheck clean, 43/43 tests pass, next build compiles
from scratch. Touched 23 source files across app/, components/, lib/.

(Several audit findings were intentionally NOT acted on after re-inspection:
the "busyRef leak" in brand/new/page.tsx was incorrect — quickAddFromGoogleInternal
owns its own busyRef lifecycle including the finally cleanup, so the reroute
path is safe. Several optional-rendering nitpicks were also skipped as low-value.)

Author: IamRamgarhia

AUDIT-PROMPT.md

@@ -0,0 +1,92 @@
+# Tool Audit Prompt — for fix-cycle agents
+
+You are auditing a specific batch of tools in the AdForge codebase. For every tool assigned to you, you must run the methodology below and produce a structured report of confirmed bugs.
+
+## Inputs you receive
+
+1. A list of tools to audit (e.g. `app/generate/meta`, `app/optimize/budget`)
+2. The companion `TEST-SCENARIOS.md` (per-tool expectations) and `AUDIT-FINDINGS.md` (recent fixes already shipped)
+3. The shared infrastructure these tools sit on:
+   - `components/GeneratorShell.tsx` — generation orchestrator
+   - `lib/llm.ts` — provider routing
+   - `lib/quota-tracker.ts` — rate-limit tracking
+   - `lib/smart-fill.ts` — brand-brain prefill heuristics
+   - `lib/storage.ts` — IndexedDB persistence
+   - `lib/brand-brain.ts` — schema + system prompt
+
+## Methodology
+
+For each tool, run these checks IN ORDER. Stop only when you've completed all six.
+
+### 1. Static read
+- Open the page file (`app/<path>/page.tsx`) and its prompt builder (`lib/prompts/<...>.ts`)
+- Read every line — no skipping
+- Confirm: does it use `GeneratorShell`? If yes, you can skip most of the rate-limit / autosave checks. If no (custom impl), audit them in detail.
+
+### 2. Happy-path trace
+Mentally simulate: user has active brand → opens tool → all fields auto-fill from brain → click Generate → output streams → saves to history.
+- Smart-fill matches: confirm the `name` of every input field in the page config matches one of the 40+ aliases in `lib/smart-fill.ts:suggestFromBrain`. List any unmatched field names.
+- buildPrompt: confirm the prompt builder produces a non-trivial prompt when given typical brain values + typical input values.
+
+### 3. Empty-input check
+Mentally simulate: user with no active brand opens the tool blank → clicks Generate.
+- Required fields gating: which fields are `required: true`? Does the UI block submission cleanly when they're missing? Or does the AI receive `undefined` and confuse itself?
+- Empty system prompt: when brain is null, `buildBrandSystemPrompt(null)` returns a generic prompt. Confirm the tool's prompt builder gracefully handles undefined brain fields (no `brain.products[0]` without null check, etc.).
+
+### 4. Streaming + autosave check (custom-implemented tools only)
+For tools that don't use GeneratorShell:
+- `const finalText = res.text || stream.text` fallback present? (audit-finding #recent)
+- `saveAd` called with the right brand_id and platform?
+- `addUsage` called with `(cost, input_tokens, output_tokens)` in the right order?
+- `estimateCostUsd` 3-arg form `(providerId, modelId, usage)`?
+- Abort signal threaded through `llmStream` / `llmCall`?
+- Rate-limit error parsing — does the catch detect 429-class errors and surface usefully?
+
+### 5. Edge cases by tool type
+- **Generator with CharBadge** (google, meta, tiktok, linkedin, etc.): does the platform-limit validation gate auto-save? Should over-limit headlines not be auto-saved? (Currently they ARE saved — flag as low-priority issue.)
+- **Generator with image field** (vision-capable): is the field config `kind: "image"`? Does the prompt include an "IF IMAGE ATTACHED" block that's conditional on actual attachment?
+- **Optimizer with budget input**: is the label currency-agnostic (no hardcoded `$`)? Does the prompt prefix the budget with the currency code from `getCurrency()`?
+- **Tool with URL input** (landing-page, reel-teardown): is `looksLikeUrl()` called? Does ingestUrl get a signal for abort?
+- **Tool with comma-separated input** (keywords): does the parse handle whitespace, trailing commas, duplicates?
+
+### 6. State-integration scenarios
+- Active brand switch mid-generation: does the tool use `runningRef` (or equivalent) to prevent brain state from mutating during an in-flight call?
+- Provider switch mid-generation: in-flight request should complete on the OLD provider; next request uses NEW.
+- Save while autosave is off: confirm the manual "Save" button is wired and works.
+
+## Output format
+
+Return a single Markdown block per tool, like this:
+
+```
+### app/generate/meta · status: PASS / ISSUES_FOUND
+
+Pass items: [list scenario IDs from TEST-SCENARIOS.md that you verified pass]
+
+Issues found:
+1. **<severity>** [file:line]: <one-sentence symptom>
+   Root cause: <what's wrong in code>
+   Fix sketch: <how to fix in 1-2 lines of code>
+   Confidence: 0-100
+
+Open questions: <anything you couldn't verify statically and would need a running browser to confirm>
+```
+
+Severity:
+- **BLOCKER**: tool crashes, silently fails, or produces wrong output
+- **HIGH**: tool works but UX is broken (no error feedback, wrong currency, etc.)
+- **MEDIUM**: edge case or nicety
+- **LOW**: nitpick, skip unless trivial
+
+## Hard rules
+
+- Do not invent issues. If you can't verify, mark it "open question."
+- Do not flag style preferences.
+- Confidence < 60 → don't include unless it's a security issue.
+- Cite file:line for every finding so the fix agent can navigate directly.
+
+## What "done" looks like
+
+- Every tool in your batch has a status block
+- Every BLOCKER and HIGH has enough detail to fix in <10 lines of code
+- The report is ready to hand to a fix agent