> **Part of [#417](#417
This PR is the **assets** — the files four conformance implementations
execute. It stacks on
[#440](#440), which configures
the releases that give these assets a tag. The normative description of
them, Appendix F, is
[#439](#439), stacked on this.
[#425](#425) adds the
machine-readable report.
>
> This PR previously carried the appendix too. It was split so the
assets can land while the prose is still being reviewed.

## What

The artifacts a provider conformance suite runs against:

| path | what it is |
|---|---|
| `assets/provider-tck/gherkin/` | six feature files — evaluation,
errors, events, lifecycle, reason, metadata |
| `assets/provider-tck/flags/canonical-flags.json` | the flag set every
scenario assumes |
| `assets/provider-tck/openapi/control-api.yaml` | the HTTP surface a
backend under test must expose |
| `assets/provider-tck/go.mod`, `embed.go` | a nested Go module exposing
the artifacts as an `embed.FS` |
| `release-please-config.json`, `.release-please-manifest.json` |
registers the assets as a releasable component; the general setup is
#440 |

Marked **experimental**: the scenario set is a representative subset,
and the capability vocabulary may still change.

## Why this is split out

Four language implementations already run these files, and all four are
currently pinned to an **unmerged branch tip**. Every rebase here
invalidates them — it has happened three times in two weeks. Nothing
about those pins becomes real until the assets are on `main` with a tag,
so landing the executable half first is worth more than landing it
together with prose that will take longer to review.

Appendix F is the same work, reviewed on its own terms.

## Why a Go module is in here

A Go module is distributed as a zip of the VCS tree, in which a git
submodule is only a gitlink and its files are absent. The other three
languages build from a working tree and keep using the submodule, but a
Go suite would have to vendor a copy of every artifact and police it
against drift. `embed.go` removes the copy: a consumer pins a revision,
the checksum database makes it immutable, and the embedded bytes are the
same bytes every other language reads out of the submodule. `go.mod` and
`embed.go` are inert for every non-Go consumer. Precedent:
`open-feature/flagd-schemas`.

The release config tags the assets as their own component,
`specification/assets/provider-tck/vX.Y.Z` — the shape `go get` needs.
Without it a Go consumer can only name a pseudo-version.

## Releasing the assets

The general release-please setup is
[#440](#440), which this stacks
on. What this PR adds is the part that describes *these* assets: a
second package, so they are tagged
`specification/assets/provider-tck/vX.Y.Z`.

That tag shape is what `go get` needs. Without it a Go consumer can only
name a pseudo-version, which is what the four conformance
implementations have had to do while this branch was unmerged, and it
has cost several rounds of re-pinning.

The specification package sets `exclude-paths` for this directory, so an
asset-only change no longer cuts a specification release, and
`separate-pull-requests` keeps the two from sharing one release pull
request.

## The capability vocabulary lives in the README, for now

A scenario exercising an optional part of the contract carries a tag; a
provider declares what it supports, and a scenario gated on an
undeclared tag is reported **skipped, with the reason** — never as
passed. A suite that quietly goes green on scenarios it did not run is
worse than no suite at all.

That rule and the tag table are in `assets/provider-tck/README.md` so
the tags in the feature files are not left undefined while the appendix
is in review. #439 moves them into the appendix and leaves a pointer
behind, so they end up with one home rather than two that can disagree.

Worth stating plainly: **withholding a tag is not an admission of a
defect.** `@numeric-coercion` borrows its rule from flagd's coercion
ADR, and `@string-typing` / `@fully-typed-values` sit on a question the
specification does not answer at all —
[#430](#430) and
[#433](#433).

## Decisions worth reviewing

- **`assets/provider-tck/` rather than beside the SDK Gherkin**, because
`assets/gherkin/evaluation.feature` already exists and tests an SDK, not
a provider. This answers #417's Q1 — the subdirectory is forced rather
than chosen.
- **`.gitattributes` normalises these to LF.** Four languages consume
them byte for byte and copy them verbatim into published artifacts, so a
Windows checkout must not package differently from a Linux one.
- **`missing-flag` must not exist** in the flag set — its absence is
what the `FLAG_NOT_FOUND` scenario tests. The other four load-bearing
properties of the flag set are in the README; each one, if seeded wrong,
turns a scenario green or red for the wrong reason in every language at
once.
- **No container restarts.** Outages are produced through the control
API, because orchestrators reassign host ports on restart and every
provider pointed at the old one then fails looking flaky.

## Implementations

All four execute exactly these files, against flagd (both resolvers),
OFREP and Flagsmith:

| | suite | flagd | OFREP | Flagsmith |
|---|---|---|---|---|
| Java |
[#1830](open-feature/java-sdk-contrib#1830) |
[#1847](open-feature/java-sdk-contrib#1847) |
[#1840](open-feature/java-sdk-contrib#1840) |
[#1849](open-feature/java-sdk-contrib#1849) |
| Go | [#940](open-feature/go-sdk-contrib#940) |
[#941](open-feature/go-sdk-contrib#941) |
[#942](open-feature/go-sdk-contrib#942) |
[#959](open-feature/go-sdk-contrib#959) |
| JavaScript |
[#1606](open-feature/js-sdk-contrib#1606) |
[#1608](open-feature/js-sdk-contrib#1608) |
[#1612](open-feature/js-sdk-contrib#1612) |
[#1623](open-feature/js-sdk-contrib#1623) |
| Python |
[#409](open-feature/python-sdk-contrib#409) |
[#411](open-feature/python-sdk-contrib#411) |
[#414](open-feature/python-sdk-contrib#414) | —
|

## What running them has found

- **flagd silently narrows a float flag to an integer**, Go and Java,
both resolvers, no error code —
[flagd#1996](open-feature/flagd#1996)
- **flagd reported ready before flags were evaluable**, on 100% of
starts —
[flagd#2047](open-feature/flagd#2047), fixed
in flagd v0.17.0
- **The Java SDK's `MultiProvider` swallows child provider events** —
[java-sdk#1882](open-feature/java-sdk#1882)
- **The Python flagd provider widens a boolean to a float** —
[python-sdk-contrib#417](open-feature/python-sdk-contrib#417)
- **The JavaScript Flagsmith provider stringifies unconditionally**
where its other paths correctly report a mismatch — visible only because
three languages ran identical scenarios against one backend
- two specification gaps, raised rather than decided here:
[#430](#430),
[#433](#433)

---------

Signed-off-by: Simon Schrottner <simon.schrottner@flagsmith.com>