# T743 — product fit ai structured fallback

> Status: `done`  
> Milestone: `M4` · Phase C  
> Kind: `feat` · Est. change: `~190 lines`

## Goal

產品適配優先使用結構化 AI 判定，輸出不完整時安全回 deterministic scorer，不保存腦補證據。

## Depends on

- T741、T742

## Inputs

- `apps/backend/internal/module/radar/usecase/judge.go`、AI provider abstraction
- `../spec.md` §3.2、§8

## Outputs

| 路徑 | 動作 | 說明 |
|------|------|------|
| `apps/backend/internal/module/radar/usecase/product_fit_ai.go` | add | prompt、JSON parse、evidence validation、fallback |
| `apps/backend/internal/module/radar/usecase/judge.go` | edit | product mode 接 fit；generic 保留舊 fit |
| AI tests | add | valid／missing dimensions／invented excerpt／provider failure |

## Behavior

- excerpt 必須可在候選文字找到；basis 必須可在 snapshot 找到。
- AI 結果缺契約即丟棄，fallback 仍需四維完整。
- 新 match 計 `ai_research`／`radar.product_fit`；既有 generic 計費不變。

## Out of scope

- merge 與 primary（T744）
- 準確度樣本（T773）

## Acceptance

- [x] provider failure 不產生空 reasons，且 token 不當會員 auth。

```bash
cd apps/backend && go test ./internal/module/radar/usecase/... -count=1
```

Implementation result (2026-08-10): product watches now attach a structured
four-dimension fit during judging and meter it as `ai_research`／
`radar.product_fit`. AI output is accepted only when every positive excerpt is
found in the candidate and every basis exists in the immutable snapshot;
malformed, invented, or provider-failed responses fall back to T742's
deterministic scorer, preserving complete reasons and hard exclusions.