phase2/negotiation_followup: $0 hold-reply never supersedes a dealer's real OTD; advisory uses the real best-competing OTD

[ryanxwy]

What

Two compounding negotiation-board correctness defects from live-e2e 2026-06-29-run2.

PIC-20260629r2-5 — a $0 hold-reply supersedes a dealer's real OTD. A "my $26,905 stands, I won't chase" reply re-extracted to a dealer_quotes row with otd_total=0; SQLite MIN()/nulls-last rank 0 as the cheapest, so the board best_otd + the latest-quote views showed $0 where a real OTD was held.

PIC-20260629r2-4 — "at or below best" advisory off by $49. The tone classifier was fed the strict itemized give-up BATNA, which dropped a cheaper bottom-line (non-itemized) competitor, so a quote $49 above the real best read "at or below best".

Fix

• persist.ts — write-normalize otd_total <= 0 → null. A no-number reply is still recorded (provenance) but never persists a $0 quote.
• migration 0006 — heal pre-existing $0 rows → NULL (data fix, no schema change; drizzle-kit generate stays no-op so db:check passes).
• followupReads.ts — latest/competing reads skip otd_total <= 0; a new bestCompetingRealOtd (lowest real same-mode competing OTD, no itemization gate) drives the tone/display, while the strict itemized BATNA still governs the give_up_switch verdict.
• negotiationProjection.ts — tone + batna_gap + best_competing_otd use the real baseline; MIN(NULLIF(otd_total,0)) heals the board roll-up.
• negotiationAdvice.ts — the give_up_switch text cites the strict itemized gap; the tone wording cites the real gap.
• giveUpProjection.ts — the board card's displayed gap mirrors the modal's real gap (verdict stays strict) so the two surfaces never disagree.

Verification

• Fresh-context code review APPROVE + safety-invariant audit SAFE + parity-oracle check PARITY OK (both defects restore parity — the pre-fix behavior was the divergence). All separate agents from the fixer.
• RUN_UI_FUNCTIONAL=1 bash scripts/green.sh → GREEN; db:check clean (drizzle-kit generate = "No schema changes"); migrator test applies 0006 cleanly.
• Deterministic regressions: persist.test.ts ($0→null at write), negotiationProjection.test.ts (board best_otd ignores $0; advisory uses the real competing OTD; non-itemized-competitor case; currentOtd ignores a newer $0), giveUpProjection.test.ts (verdict strict, displayed gap real), negotiationAdvice.test.ts (give_up_switch cites the itemized gap, not the tone gap).
• Paired LLM-seasoned cases registered (re-quotes-then-holds dealer; non-itemized cheaper front-runner).