Round-4 money audit: --max-spend ceiling + paid-tool accounting + trading guards (3.29.8)

[VickyXAI]

Fourth review round, sweeping the paid-tool surface beyond the diff-scoped rounds 1–3. 13 adversarially-verified findings + a follow-up on the just-merged #89. All money-critical; local suite 479/479.

Wallet spend ceiling (HIGH)

• --max-spend was blind to ALL paid-tool USDC (ImageGen/VideoGen/MusicGen, Exa, Surf, RealFace, Voice, Phone, Modal, DeFiLlama, RPC, Prediction) — only LLM token cost counted, so franklin -p "generate 30 images" --max-spend 0.50 could spend multiples with the guard never firing. New getLiveSpendUsd accumulator is diffed around tool execution and folded into the session total; ceiling re-checked after each tool batch.
• Charged-but-failed LLM calls now counted. A paid call whose stream then failed (timeout / mid-stream error / Esc) debited USDC but was never folded into the cap, and drift compounded across --resume. The loop catch captures getLastPaidUsd() before the next attempt resets it.

Paid-but-invisible accounting

• DeFiLlama (5), MultiChainRPC, PredictionMarket (now persisted), MusicGen, panel phone Buy/Renew/List now recordUsage the settled amount (parity with surf.ts) — were charging real USDC invisible to franklin stats (reconciliation couldn't even detect the gap).

Money-safety + correctness

• $5 BuyPhoneNumber/RenewPhoneNumber now confirm via onAskUser (skipped under FRANKLIN_MEDIA_AUTO_APPROVE_ALL) — were ungated while a $0.05 Modal create gates.
• VoiceStatus records one row per invocation, not per 5s poll (was up to 420/call, evicting real spend history at the 1000-row cap).
• RiskEngine per-position cap values the held position at cost basis, not the order price (a buy could slip the cap after a price drop).
• closePosition clamps an over-sized close to the held qty instead of throwing only X held.

PR #89 follow-up (thanks @samsamtrum)

Merged #89 (reject sub-precision Jupiter amounts that rounded up to one atomic unit), then refined: floor excess precision to the atomic unit (so agent-computed amounts / float noise still swap) and reject only true dust, with a clearer message + unit test.

New no-spend regression tests: live-spend ceiling basis, risk cost-basis cap, close clamp, Jupiter atomic-unit math.

🤖 Generated with Claude Code

[VickyXAI]

Independent review — 1 real bug + 1 nit found & fixed (928cfea)

Ran a 4-lens adversarial review of this PR before merge. It caught a real bug in the headline change:

• MEDIUM (fixed): --max-spend missed concurrent paid tools. DeFiLlama/RPC/Prediction are concurrent: true and start mid-stream via onToolReceived, so they could settle their x402 charge before the post-stream getLiveSpendUsd() snapshot — baking their spend into the baseline and excluding it from the cap. Two lenses flagged this independently. Fix: snapshot before the stream and net out the LLM's own callCost, so the fold captures concurrent + sequential tools alike.
• Nit (fixed): toAtomicUnits scientific-notation path rounded (9.999e-10 → 1 atomic unit — the dust round-up Reject Jupiter amounts below token precision #89 killed) and threw a raw BigInt error on ≥1e21. Now uses toLocaleString (plain decimal, floor-by-truncation) so it floors and rejects only true dust on all inputs. Added boundary tests.

One finding (the ceiling-basis test "gives false confidence") was correctly rejected as a coverage nit, not a defect — though it points at the same concurrent-tool gap, now closed.

Local suite 479/479. Ready for merge review.