[FINAL ] Sample App

### Motivation

Follow-up test coverage for #3603. That PR's refactor commit (`b2bfd86`)
fixed a bug where `workflowId`/`stepId` were `@Transient` on the cached
transaction metadata and were lost on the disk round-trip, so
`presented_workflow_id` / `presented_step_id` would be sent as `null` on
retry or unsynced replay. The field was promoted to a first-class
`workflowMetadata` on `LocalTransactionMetadata` with
`@SerialName("workflow_metadata")`, but there was no regression test
asserting it actually survives serialization.

### What this adds

One test in `LocalTransactionMetadataStoreTest`,
`cacheLocalTransactionMetadata round-trips workflow metadata`, mirroring
the existing `handles paywall data` test. The JSON written to
`SharedPreferences` is captured and fed back to
`getLocalTransactionMetadata`, so the real `JsonTools.json`
encode/decode runs (only SharedPreferences I/O is mocked).

### Verification (RED → GREEN)

- **RED**: with `workflowMetadata` marked `@Transient`, the test fails
at the assertion (`retrieved?.workflowMetadata` is `null`).
- **GREEN**: with `@SerialName("workflow_metadata")` (current state), it
passes.

Ran `:purchases:testDefaultsBc8DebugUnitTest --tests
"...LocalTransactionMetadataStoreTest"` locally (JDK 21).

Companion to #3636 (which covers the `PostReceiptHelper` threading and
replay paths). Together they close the two test gaps identified in
review of #3603.

<details>
<summary>AI session context</summary>

**Task:** Review #3603 ("is anything missing?"), then close the
identified test gaps.

**Findings from review of #3603:**
1. (important) Disk round-trip of `workflowMetadata` untested, the exact
bug `b2bfd86` fixed. `LocalTransactionMetadataStoreTest` had the
analogous `handles paywall data` round-trip test but nothing for
`workflowMetadata`.
2. (important) `PostReceiptHelper` threading untested (only an `any()`
count bumped). → covered by #3636.

**This PR** addresses finding #1.

**Decisions:**
- Reused the existing capture-the-written-JSON pattern so the test
exercises real serialization, not a mock, which is what makes it a
genuine guard for the `@Transient` regression.
- Based on `cesar/receipt-presented-paywall-workflow-ids` because the
`workflowMetadata` field only exists there, not on `main`.

**Verified:** RED→GREEN as described above. Production code
(`LocalTransactionMetadata.kt`) left unchanged; only the test file is
added.

**Not done:** The remaining review items on #3603 were questions, not
test gaps (all-or-nothing `WorkflowMetadata.from`, impression/purchase
step-id divergence under dedup, `pr:other` vs `pr:fix` label). Not
addressed here.

🤖 Generated with [Claude Code](https://claude.com/claude-code)
</details>

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>