feat: greenfield spec-test synthesis for dbt-pr-review (raise the no-reference verification ceiling)

## Problem

`dbt-pr-review` verifies **modified** models via base-vs-head equivalence, but for **greenfield** (git-added) models there is no reference, so the reviewer can only lint them — the verification ceiling is \"it builds + whatever tests the author happened to write.\" `missingGrainTestLane` detects the gap but does nothing about it.

## Approach (BAIR \"Intelligence is Free\" — data systems *by* agents)

Auxiliary agent that generates tests from a new model's **declared intent** to expand the spec, in two strictly separated tracks:

- **Track A (deterministic, block-eligible):** materialize dbt constraints the author *declared but did not enforce* (`schema.yml` `not_null`/`unique`/`accepted_values`/`relationships`, contract types). Assertion built from the parsed constraint — **no LLM in the blocking path**, so it can't be tricked into a back-labeled false positive.
- **Track B (LLM, advisory):** propose *new* tests from soft intent (descriptions, `ref()` edges, PR text). Emitted at `confidence: unknown`, structurally excluded from the verdict — never blocks.

Specs: `packages/opencode/specs/greenfield-spec-test-synthesis.md` (+ `corrective-app-memory.md` for the P2 feedback loop).

## Phasing

- **P0 (this issue's first PR):** propose-only — generate candidate tests + a committable `schema.yml` patch, all non-blocking. Authoring help, not \"verification.\"
- **P1:** `declaredConstraints(model)` runner API + execution + the one deterministic blocking path + verdict-layer enforcement + assertion sandbox.
- **P2:** corrective per-repo memory feeding priors back into generation.