Say upstream's live tests pass on the mlx backend, list each package product's targets, bring CONTRIBUTING up to 0.1.0, and restore the README's read-speed bullet

[alaineid]

Refs #65.

Fixes the stale documentation lines found while preparing release 0.1.0 (#145), and restores a
README line that #145's merge with #144 dropped. Docs only: no Swift code, configuration or test
changes. CI does not run, because every changed file is Markdown that no test reads.

docs/development.md, The live suite (line 317)

Was: "Against the Swift server's mlx backend, its think, chat and stream tests fail until #52
and #53 land, while the Swift suite skips them. The chat tests run there rather than skip because
the listing already names diffusiongemma-26b, as upstream's does."

Now:

• Upstream's file runs 12 tests there and skips the four test_encoder cases.
• All 12 passed on 2026-10-06, test_think, test_chat and test_chat_stream among them (The think option: thought generation before a read, prefix continuation, billing #52
  and /v1/chat/completions: OpenAI-compatible generation with streaming, JSON mode and cancellation #53 landed in Generate text, think and chat on the MLX backend: the diffusion sampler, the block loop, think and the chat model (#50, #51, #52, #53, D-059) #137, D-059).
• The chat tests run because the listing names diffusiongemma-26b.
• The Swift suite runs the same tests on the same conditions and passed them too, with its
  long-stream check.

Evidence: a run on 2026-10-06 at bbab174 (#137), on an M3 Max with macOS 27.0.1 and Swift 6.4,
following the section's own steps. Since then main has changed only the version strings (#146)
and DocC Markdown (#147), so the result holds for d176ce3.

• swift build -c release --product openjev, then
  OPENJEV_MLX_CACHE_LIMIT_GB=4 OPENJEV_HOST=127.0.0.1 OPENJEV_PORT=8093 openjev serve --backend mlx.
  It loaded the default checkpoint, mlx-community/diffusiongemma-26B-A4B-it-4bit at a7a81407,
  from the Hugging Face cache. /v1/models listed openjev-latest, openjev-0.1 and
  diffusiongemma-26b.
• make upstream (dcd2094) and Upstream/.venv with pytest 9.1.1 and httpx 0.28.1, then
  OPENJEV_LIVE_URL=http://127.0.0.1:8093 Upstream/.venv/bin/python -m pytest Upstream/openjev/tests/test_live.py -v:
  12 passed, 4 skipped in 60.95s.
• The Swift suite: the OpenJevLiveTests bundle that swift test --filter OpenJevLiveTests runs.
  It was built with swift build --target OpenJevLiveTests and run with swiftpm-testing-helper,
  so no other test target had to build.
  • Its test_live.py suite passed 13: the same 12 plus "A long streamed reply arrives whole,
    with its finish and [DONE], as the reply unstreamed".
  • It skipped the 4 test_encoder cases ("the server at OPENJEV_LIVE_URL does not list ...").
  • The run ended "Test run with 28 tests in 3 suites passed after 90.915 seconds."
• Free disk stayed at 46 GB and swap at 162 MB, and the server stopped cleanly on SIGINT.

pytest -v

test_live.py::test_readme_example PASSED
test_live.py::test_image PASSED
test_live.py::test_read_options[extra0] PASSED
test_live.py::test_read_options[extra1] PASSED
test_live.py::test_read_options[extra2] PASSED
test_live.py::test_think PASSED
test_live.py::test_255_options PASSED
test_live.py::test_many_questions_in_chunks PASSED
test_live.py::test_unknown_model PASSED
test_live.py::test_concurrent_reads PASSED
test_live.py::test_chat PASSED
test_live.py::test_chat_stream PASSED
test_live.py::test_encoder[laya-1.0] SKIPPED
test_live.py::test_encoder[verdict-1.4] SKIPPED
test_live.py::test_encoder[clm-v0.1] SKIPPED
test_live.py::test_encoder[jevk5-0.2] SKIPPED
=================== 12 passed, 4 skipped in 60.95s (0:01:00) ===================

docs/development.md, Dependencies (lines 99 to 106)

The "Products used" column now lists every product Package.swift uses, each with the targets
that link it. It keeps the server rows' names (server, CLI, stub server) and adds DiffusionGemma,
encoders, JevK5 and bench, the names the Targets bullets above the table use. Test targets are
spelled out one by one ("server tests, CLI tests"), because a grouped "server, CLI and JevK5
tests" can be read two ways. A script parses the column back into target names and compares it
with every target's .product(name:package:) entries in Package.swift: 22 products in each, no
mismatch.

Row	Was	Missing or wrong against Package.swift
mlx-swift	MLX, MLXNN	no targets named
mlx-swift-lm	MLXLMCommon, MLXVLM	MLXLLM (OpenJevLetterReadout) and MLXHuggingFace (OpenJevDiffusionGemmaTests) missing; MLXVLM is linked only by OpenJevDiffusionGemmaTests (D-054)
swift-transformers	Tokenizers	Hub (OpenJevDiffusionGemma, OpenJevLetterReadout) missing
swift-jinja	Jinja	no target named
hummingbird	Hummingbird, HummingbirdCore (server), HummingbirdTesting (server and CLI tests)	Hummingbird is also linked by OpenJevCLITests and OpenJevLetterReadoutTests, and HummingbirdTesting by OpenJevLetterReadoutTests
swift-argument-parser	ArgumentParser	no targets named (openjev, openjev-bench and their tests)
swift-http-types	HTTPTypes (server, server and CLI tests)	also linked by OpenJevLetterReadoutTests
swift-log	Logging (server, CLI, server and CLI tests)	also linked by openjev-stub-server

The swift-nio, swift-service-lifecycle, async-http-client and swift-docc-plugin rows already
matched and are unchanged. The Requirement and Resolved columns match Package.swift and
Package.resolved: 35 pins, these 12 and 23 transitive, as the text under the table says.

docs/05-architecture.md (lines 37 to 39, 70, and 132 to 135)

These are the same product facts, in the module tree and graph whose product names development.md
says it matches.

• Was. The graph still gave OpenJevDiffusionGemma mlx-swift-lm (MLXLMCommon, MLXVLM),
  though D-054 removed MLXVLM from the library (the tree in the same file says "not MLXVLM").
  Neither the tree nor the graph named Hub (import Hub in SwiftTransformersTokenizer.swift
  and JevK5Tokenizer.swift) or swift-jinja (import Jinja in SwiftTransformersTokenizer.swift).
• Now. OpenJevDiffusionGemma depends on mlx-swift, MLXLMCommon, swift-transformers'
  Tokenizers and Hub, and swift-jinja. OpenJevLetterReadout also depends on Hub.

CONTRIBUTING.md (lines 3 to 11)

• Was. "The project is in its research and planning phase."
• Now, the state. Milestones 0 to 5 are complete, and release 0.1.0 is the first version a
  package can depend on, which is Prepare release 0.1.0: versioning policy, changelog, security notes, Swift Package Index settings, code owners, issue templates, adopters and release notes (#65) #145's README Status wording. The opening then points to the
  README's Status section for what is done and what is not there yet, rather than repeating that
  list. The milestones API reports milestones 0 to 5 closed with no open issue ("5. Text
  generation and think": 0 open, 5 closed).
• Now, the pointers. Prepare release 0.1.0: versioning policy, changelog, security notes, Swift Package Index settings, code owners, issue templates, adopters and release notes (#65) #145 has merged, so a second paragraph points at the issue templates in
  .github/ISSUE_TEMPLATE/, described as their own description fields describe them (a bug
  report and a compatibility report), and at SECURITY.md for vulnerabilities, as the templates'
  config.yml does.

README.md, Status (line 96)

• Restored. "Faster DiffusionGemma reads: Expert matmuls take 61% of a read: measure the sort threshold, the gathered path and a compiled expert block #100, Long-prompt prefill: the rate falls from 1,300 to 760 tokens/s between 1,000 and 10,000 tokens #101 and Throughput stays at 3.0 to 3.4 requests/s from 1 to 16 concurrent reads: batch the decoder passes #102" replaces "Release 0.1.0, with a
  version tag a package can depend on: Release 0.1.0: versioning, tag, Swift Package Index, changelog, security notes #65" under "Not there yet". That is Prepare release 0.1.0: versioning policy, changelog, security notes, Swift Package Index settings, code owners, issue templates, adopters and release notes (#65) #145's own change,
  which its merge of main lost.
• How it was lost. Defer CLM until someone asks for it: a D-011 addendum, and the pages that said it was coming (#59) #144 (CLM deferred) merged at 14:58 and rewrote the same list. The merge
  1daa94a ("Merge branch 'main' into release-0.1.0-prep") then resolved the conflict by taking
  Defer CLM until someone asks for it: a D-011 addendum, and the pages that said it was coming (#59) #144's side whole (git show --remerge-diff 1daa94a -- README.md), so Prepare release 0.1.0: versioning policy, changelog, security notes, Swift Package Index settings, code owners, issue templates, adopters and release notes (#65) #145's squash faf33c0
  never changes the list.
• Why it matters. Main's README now says "Release 0.1.0 is the first version a package can
  depend on" in the paragraph above and lists it as not there yet. Expert matmuls take 61% of a read: measure the sort threshold, the gathered path and a compiled expert block #100, Long-prompt prefill: the rate falls from 1,300 to 760 tokens/s between 1,000 and 10,000 tokens #101 and Throughput stays at 3.0 to 3.4 requests/s from 1 to 16 concurrent reads: batch the decoder passes #102 are open
  and are about read speed.

Not changed: docs/development.md lines 89 and 90 (a question)

The lines say: "A package opens in Xcode with autogenerated schemes, which are per-user and are
never written to disk, so this one is committed."

• What is in the repository. Since Read images on the MLX backend: the Gemma 4 vision tower, the image prefill and the images field (#47, #48, D-054) #129 (046a26f), eight autogenerated schemes are committed
  beside OpenJevCore-iOS.xcscheme in .swiftpm/xcode/xcshareddata/xcschemes/: OpenJevCore,
  OpenJevDiffusionGemma, OpenJevEncoders, OpenJevLetterReadout, OpenJevSwift-Package, openjev,
  openjev-bench and openjev-stub-server. .gitignore lets any *.xcscheme there through.
• Why it is left as it is. Either the schemes are deleted (and the exception narrowed to
  OpenJevCore-iOS.xcscheme), or they stay and the lines are reworded. That is the maintainer's
  call.

Merge order

The 0.1.0 tag does not exist yet. The README's Status paragraph (#145), this PR's CONTRIBUTING
opening and the restored bullet all describe the release as published, so they read right once
0.1.0 is tagged.

Sweep

Every tracked text file was searched with comment markers stripped, lines joined and the text split
into sentences. This was done again after rebasing onto #144 to #147.

• No other copy of "planning phase".
• The sentences that put think or chat behind The think option: thought generation before a read, prefix continuation, billing #52 or /v1/chat/completions: OpenAI-compatible generation with streaming, JSON mode and cancellation #53 are in two places, both left as written:
  • docs/06-decisions.md, which records what was true at the time.
  • docs/09-conformance-and-testing.md's dated run history, which already records the /v1/chat/completions: OpenAI-compatible generation with streaming, JSON mode and cancellation #53 run.
    That record gives the Swift suite 12 passes with chat. It was written in 265ebbb, before
    Generate text, think and chat on the MLX backend: the diffusion sampler, the block loop, think and the chat model (#50, #51, #52, #53, D-059) #137's review added the long-stream test (0b771b8), so today's 13 is that test, not a
    change in the others.

[Copilot]

Copilot review overview

🟢 Approval recommended

The documentation-only changes align with the supplied implementation and release evidence, with no unresolved findings.

Review effort: Balanced
Findings: None

What changed in this PR

Updates release 0.1.0 documentation to reflect shipped functionality and current dependencies.

Changes:

• Records passing MLX live tests, including think and chat.
• Corrects dependency product and target documentation.
• Refreshes contributor guidance and restores the pending read-speed improvements.

File	Description
README.md	Restores the read-speed improvement bullet.
docs/​development.md	Updates dependency mappings and live-test results.
docs/​05-architecture.md	Corrects module dependency descriptions and graph.
CONTRIBUTING.md	Updates release status and issue-reporting guidance.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.