Instrument Android checkout failures

[tiagocandido]

Summary

• Record Android checkout errors, protocol decoding failures, paired navigation retry outcomes, and navigation duration
• Add a default-on telemetry setting with an explicit runtime opt-out that discards buffered measurements
• Lazily and failure-safely initialize telemetry, avoiding threads and wake-ups on the disabled path, and guard recorder creation against a concurrent runtime opt-out
• Report react-native-android as the platform when embedded in React Native, android otherwise
• Keep retry reasons paired across attempts and preserve preload attribution captured at navigation start
• Match the cross-platform decode taxonomy: envelope decode failures record envelope and params decode failures record params
• Record decode errors once per undecodable message, not once per client decode attempt: a message decoded by both a merchant handler and the kit default client counts once, and a terminal ec.error with undecodable params records even when nothing subscribes to it
• Document the bounded diagnostic data and explicit opt-out

Validation

• Android unit tests
• Detekt
• Android API dump and API compatibility checks

[tiagocandido]

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

• Instrument web checkout failures #678
• Complete React Native telemetry opt-out #677
• Instrument Android checkout failures #676 👈 (View in Graphite)
• Instrument Swift checkout failures #675
• Add internal Checkout Kit telemetry foundations #674
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[github-actions]

Package Size

Platform	Artifact	Base	Head	Delta
Android	release AAR	349.8 KiB	364.6 KiB	+14.8 KiB

Android file breakdown

File	Base	Head	Delta
classes.jar	372.3 KiB	388.2 KiB	+15.9 KiB
res/layout/checkout_view_content.xml	2.6 KiB	2.6 KiB	0 B
res/layout/checkout_sheet_content.xml	2.0 KiB	2.0 KiB	0 B
res/values/values.xml	1.3 KiB	1.3 KiB	0 B
R.txt	1.2 KiB	1.2 KiB	0 B
AndroidManifest.xml	922 B	922 B	0 B
proguard.txt	798 B	798 B	0 B
res/drawable/close.xml	431 B	431 B	0 B
res/menu/checkout_menu.xml	354 B	354 B	0 B
META-INF/com/android/build/gradle/aar-metadata.properties	157 B	157 B	0 B

Measured from the PR base SHA and PR head SHA. The file breakdown shows uncompressed sizes within each package artifact, so individual files do not sum to the compressed artifact total. This comment reports package artifact sizes only; it is not a final app binary-size report.

[bitrise]

Install this build

Open Tophat, select your target device, then click Install. Links open on the Mac running Tophat.

SDK	Install
Kotlin	Install with Tophat

Checkout Kit E2E results

Status	Suite	Target	Platform	OS version tag	Device
✅	kotlin-android	kotlin	android	latest	Google Pixel 11 Android 17.0

[adrian-augustyn]

The PR description mentions matching the cross-platform decode taxonomy, which made me wonder whether the counts match across platforms too, not just the labels. an AI review pass flagged this spot, and I traced it through the diffs to check.

Since this hook is built into every CheckoutProtocol.Client, the count seems to depend on how many clients process a message, rather than on the inbound message itself. Two cases I wanted to sanity-check:

• if a merchant registers an ec.complete handler, AlwaysRunAfterMerchant means a malformed payload can be decoded by both the merchant client and the kit default client, so it looks like it records twice
• if ec.error arrives with a valid envelope but malformed params and there's no merchant subscriber, it looks like no handler decodes it, and the terminal-error path decodes it separately without recording a params decode error, so it records zero

So the question is: is this metric meant to count failed decode attempts by clients (which is what it does today), or malformed messages from checkout (which is what the name suggests to me)? If it's the latter, should it record once per message in EmbeddedCheckoutProtocolBridge, where each message is first decoded?

[tiagocandido]

You read it right — the contract says undecodable messages, so per-client counting was wrong on both edges. The bridge now records once per message, and the terminal ec.error path records even with no subscribers; both have regression tests. iOS had the same ec.error gap (fixed in #675), web already counts per message.