[leoromanovsky]

What does this PR do?

Make native flag evaluations and emitted events conform to the merged shared contract. The 63-case snapshot pins fixture main at 9afc3f75c3fef9643a9c1b5e2b1e5a9adac48177 and runs through the public client before and after cache restoration. Ref: FFL-3568.

Motivation

Consumers need one tested rule for numeric reads, unsupported variation types, obfuscated flag keys, and event aggregation. A typed read must not silently lose integer precision or log a successful exposure for a rejected conversion.

Additional Notes

Stacked on #4099, retaining its converter structure and bounded decimal parsing. The additional checks align it with the merged contract: fractional Double reads round, whole values must be exactly representable, and overflow/nonzero underflow return TYPE_MISMATCH. For example, 1.20000000000000001 returns 1.2; 1e300 is rejected as an inexact whole-number Double read. Malformed numeric text remains a parse error.

Unsupported variation types are omitted with a once-only internal user warning, leaving valid siblings available. Missing extraLogging defaults are compatible with shared responses. Obfuscated lookups and real emitted records retain plaintext flag identity. The runner reports an actionable fixture error for an applicable null targeting key; the native context API requires a string.

The shared importer preserves provenance and response tokens. A weekly workflow prepares draft fixture updates. Obsolete converter detekt entries were removed; entries required by the merged code remain.

Validation: 1,007 feature unit tests and 259 integration tests pass, with 3 integration platform exclusions. Affected-module formatting and detekt pass. All snapshot integrity/source checks pass. Hosted comparison is tracked in the native preview rerun; these are unit-level checks, not deployed-backend validation.

Review checklist (to be filled by reviewers)

• Feature or bugfix MUST have appropriate tests (unit, integration, e2e)
• Make sure you discussed the feature or bugfix with the maintaining team in an Issue
• Make sure each commit and the PR mention the Issue number (cf the CONTRIBUTING doc)

[datadog-datadog-us1-prod]

 

🎉 All green!

🧪 All tests passed
❄️ No new flaky tests detected

🎯 Code Coverage (details)
• Patch Coverage: 95.61%
• Overall Coverage: 72.18% (+0.13%)

_{This comment will be updated automatically if new data arrives.
🔗 Commit SHA: 96b2bf6 | Docs | View more details | Give us feedback!}

[aarsilv]

🤖 Comment from Claude, working with Aaron:

This PR and #4099 both change FlagValueConverter. Could you stack this PR on #4099 and keep its converter? This PR would then keep the fixture runner, the sync workflow, the extraLogging default and the handling of unsupported variation types. If you prefer your rules for the two differences below, we can change #4099 first.

#4099 now checks whole values in Double reads as well as Int reads. We ran the shared fixtures from DataDog/ffe-system-test-data#65 against it with this PR's runner, merged with develop. The local test copy adds "extraLogging": {} to each fixture assignment. Of 95 runs, 88 pass, 3 are skipped and 4 fail. The 4 failures expect FLAG_NOT_FOUND for an unsupported variationType, and this PR omits those flags.

Two differences to settle:

• FFL-3565: Accept every numeric type in Int and Double reads #4099 sends at most 64 characters to BigDecimal. This PR has no length limit. On the JVM, stripTrailingZeros() took about 10 s for 1 followed by 200,000 zeros and .0.
• FFL-3565: Accept every numeric type in Int and Double reads #4099 reads an over-precise decimal such as 1.20000000000000001 as the nearest Double. This PR returns TYPE_MISMATCH. A review comment on ffe-system-test-data#65 asks the fixture author for a rule.

Could you also update the description's statement that #4099 "still enforces the previous strict-type contract"? #4099 already has lossless Int reads across all numeric types.

[aarsilv]

🤖 Comment from Claude, working with Aaron:

P3 follow-ups after rebasing on #4099:

• README numeric text. The new section says that inexact conversions return TYPE_MISMATCH. With FFL-3565: Accept every numeric type in Int and Double reads #4099, a Double read of 1.20000000000000001 returns the nearest Double. For NaN, resolve() reports PARSE_ERROR. Could this section link to the existing numeric paragraph in the README?
• detekt entries. Most new entries support this PR's converter. Examples include BigDecimal.intValueExact(), Result.success(kotlin.Int), and Double.isFinite(). After rebasing, could you remove entries that the resulting code no longer uses? Keep entries for calls that remain, including EnumEntries.none in FlagsJson.
• Unsupported types. PrecomputeMapper drops flags with an unsupported variationType without logging the cause. Once the client is ready, a lookup reports FLAG_NOT_FOUND, as the fixtures expect. Could the mapper log a warning once through the internal logger to explain the cause?
• targetingKey: null. The fixture schema permits context.targetingKey to be null. At line 63, CanonicalPrecomputedTest calls context["targetingKey"].asString, which throws for JSON null. None of the 57 current fixtures uses a null targeting key. Could the runner handle null, or could future fixtures with null exclude Android through _skip?

[leoromanovsky]

@aarsilv Follow-ups addressed in a45b99c342 and the numeric documentation follow-up:

• Stacked on FFL-3565: Accept every numeric type in Int and Double reads #4099 with a merge commit and changed the PR base. Retained its converter structure and 64-character decimal parsing bound. The merged fixture contract requires nearest-Double fractional reads, exact original whole-number reads, and rejection of nonzero underflow. The remaining delta enforces those rules, including rejecting 1e300 as an inexact whole value. The PR description and README no longer describe FFL-3565: Accept every numeric type in Int and Double reads #4099 as strict-type or promise the old printed-form exception.
• Removed obsolete converter detekt entries; retained the calls the merged implementation uses, including the unsupported-type detection.
• Added a once-only internal USER warning for omitted unsupported variation types, plus coverage that valid sibling flags survive and the warning requests onlyOnce.
• Defined null-targeting-key applicability explicitly: Android requires a string, so an applicable null case fails with an instruction to declare the exclusion upstream instead of throwing from asString.

The snapshot pins fixture main at 9afc3f75. Local validation passes: 1,007 feature tests, 259 integration tests with 3 platform exclusions, formatting, and detekt.