Cloud sweep: fix ch_func_to_all_pk with *args, partialx reordering, pyproject + uv CI, runnable README

[thorwhalen]

Tests

Command (from the repo root):

uv venv .venv && uv pip install -e ".[dev]"
.venv/bin/python -m pytest          # pyproject addopts: --doctest-modules, ignores i2/examples and i2/scrap

	passed	failed	xfailed
Baseline (master 14bd29d)	758	0	2
Final (this branch)	785	0	2

The CI-equivalent command also gives 785 passed: --cov=i2 --doctest-modules -o doctest_optionflags='ELLIPSIS IGNORE_EXCEPTION_DETAIL'. The two fix commits alone, without the packaging commit, give 783 passed on the legacy setup. The new uv CI is green on 3.10 and 3.12 for the current head.

Dependents gate: each dependent was cloned from GitHub and run with i2 installed editable, first from master, then from this branch.

Dependent	Baseline	Final
config2py	132 passed	132 passed
meshed	204 passed, 1 failed	204 passed, 1 failed
ju	46 passed	46 passed
dol	592 passed	592 passed

meshed's one failure is test_hybrid_dag. It fails the same way on master because the optional deps http2py and extrude are missing.

Changes

• fix: make ch_func_to_all_pk work for functions with *args: on master, every call failed or lost arguments when the function had *args. Only that case takes the new path. Functions without *args keep master's code path and its convention: **kwargs given by name as a dict, other extras ignored. meshed's hook_up and FlexFuncFanout rely on that convention, and tests now pin it. Fixes Make ch_func_to_all_pk work with kwargs #12.
• fix: keep parameter kinds in order when partialx reorders params: move_params_to_the_end now stable-sorts by parameter kind. Only orders that used to raise change. Fixes Careful reordering when adding keyword-only params to a list that already has a variadic keyword. #17.
• build: migrate packaging to pyproject.toml and CI to wads uv: hatchling pyproject.toml with SPDX license and full metadata, plus the wads inline uv workflow. It is inline because the reusable stub can't pass the epythet v2 Pages pin, which is kept. setup.cfg, setup.py and the stale i2/tests/test_requirements.txt are removed. This commit is self-contained and can be dropped on its own.
• docs: fix broken README examples and guard them with a test: seven README examples didn't run and now do. i2/tests/test_readme.py runs every README python block. The README also gains an agent pointer and a "For carbon-based contributors" section, and .claude/CLAUDE.md is updated. Fixes Seven README examples don't run #94.

Issues

• Fixed here: Make ch_func_to_all_pk work with kwargs #12, Careful reordering when adding keyword-only params to a list that already has a variadic keyword. #17, Seven README examples don't run #94. Seven README examples don't run #94 was filed this session.
• Filed: tuple_the_args: passing the tupled *args by keyword fails, and ch_func_to_all_pk can't handle a required keyword-only after *args #93, two follow-ups found while fixing Make ch_func_to_all_pk work with kwargs #12:
  • tuple_the_args rejects its own tupled param when it's given by keyword.
  • ch_func_to_all_pk can't handle a required keyword-only param after *args.
    Both need a design call and are labelled needs-local.
• Commented: Change replace_kwargs_using so it will create keyword-only params for wrapped variadic #78. The keyword-only behaviour it asks for already landed in f84b5a9 (2025-10-11). The remaining gap is that a source *args becomes a required args, and fixing it isn't purely additive. Labelled needs-local.
• Labelled needs-local: Need a more powerful and clean signature merging functionality #43, the signature-merging design; I created the label.
• Closed: none. Nothing met the bar of evidence for closing.

Not done, and why

• Design threads on Sig and signature merging (Write a general parameters merging function #5 through Add attribute assignment dependencies in attribute_dependencies #75, Need a more powerful and clean signature merging functionality #43, Using a NotSet sentinel to solve make FuncFactory instances' signature more truthful about argument requirements. #48): out of scope, since they need Thor's design decisions.
• Reusable-workflow stub: kept inline to preserve the epythet v2 pilot pin. Once the pin is dropped, wads-migrate ci-to-stub applies.
• Skills layout: skills were not moved to i2/data/skills with .claude/skills symlinks, the skill-enable layout. That would change the wheel contents, so it's left for a local pass. The README's gh skill install ... --allow-hidden-dirs line comes from the house docs and was not run here, because there's no gh in the VM.
• ruff format: not applied. The first master publish will reformat 9 files in the bot commit. The reviewer confirmed the suite passes on the formatted tree.
• Tooling note: wads-migrate ci-to-uv carried over the dead SCRIPTS_REPOSITORY_URL with a wrong alias to SCRIPTS_USERNAME. I removed it by hand. That's worth fixing in wads.

Adversarial review (Opus subagent)

The final verdict was APPROVE after three rounds.

• Round 1, BLOCK. The first version of the Make ch_func_to_all_pk work with kwargs #12 fix changed behaviour for functions with **kwargs but no *args. It nested kw={...} instead of spreading it and passed extras through instead of dropping them. That broke FlexFuncFanout kwonly calls and changed meshed hook_up output, which meshed's tests don't cover. Fixed: master's code path is restored for functions without *args, and tests pin it. Re-running the reviewer's scripts against master: every remaining difference is a call that raised on master.
• Round 1, non-blocking findings:
  • The Careful reordering when adding keyword-only params to a list that already has a variadic keyword. #17 test didn't assert the resulting signature. It now asserts it, and the docstring documents that names only move within their kind.
  • doctest_optionflags differed from CI. It is now aligned.
  • ruff format would change 9 files. Left as is; see above.
  • Publishing now assumes the PYPI_PASSWORD secret is a PyPI API token. See below.
  • The wheel also ships i2/tests/test_readme.py, which skips itself when the README is absent. Otherwise the wheel is file-for-file identical to the setuptools build.
• Round 2, APPROVE. Non-blocking: the returned __signature__.name had become 'Sig' instead of the function's name. Fixed and tested.
• Round 3, APPROVE. Non-blocking: two edge cases in the new "**kwargs given by name" spreading, both for calls that raised on master.
  • A dict key named like the *args param is taken as the positional args.
  • A clash between the dict and an explicit keyword resolves to the explicit keyword.
    Both are left as is.

For whoever lands it

• Left open on purpose: I didn't self-merge. The packaging commit replaces the publish pipeline, and its publish and Pages jobs only run on master, so they can't be exercised from a PR. The triage note also asked that it stay droppable for a local lander.
• Merging all four commits:
  • The first master run publishes through uv publish with secrets.PYPI_PASSWORD as a token. The 0.1.74 upload succeeded, so it is very likely already a token. If it isn't, publish fails before the bump and tag, so nothing half-publishes.
  • The epythet v2 Pages pin is preserved.
• Dropping the packaging commit: the other three apply cleanly on the legacy setup (783 passed).
• Version: unchanged at 0.1.74. CI bumps it.

Fixes #12
Fixes #17
Fixes #94

🤖 Generated with Claude Code

[thorwhalen]

cloud-status: started — baseline 758/0 failed/2 xfailed; plan: fix #12 ch_func_to_all_pk kwargs (done), fix #17 partialx reordering, #78 if additive, packaging migration own commit, dependents gate (config2py, meshed, ju, dol)

---

Generated by Claude Code

[thorwhalen]

cloud-status: done — tests 758 → 785 passed (2 xfailed); dependents unchanged; Opus review APPROVE; CI green; fixes #12, #17, #94 (filed); filed #93; needs-local #43, #78, #93; left for local landing (includes the pyproject + uv CI migration, so its publish path is untested)

---

Generated by Claude Code

[thorwhalen]

Review from the local session that decided #78 and #93 (with a use-site inventory across i2's dependents and a test sweep).

Verdict on this PR's approach: confirmed, and built on. The *args path of ch_func_to_all_pk and the _args_and_kwargs_from_all_pk_arguments helper are kept exactly as they are; tuple_the_args now uses the same helper (#93 part 1), so the two functions share one rebuild path. The one thing this PR left as "not yet handled", a required keyword-only after *args, is handled in all_pk_signature (#93 part 2) and needed no change to this PR's call path. #78 is resolved by dropping a source *args in replace_kwargs_using. The decisions and evidence are on the two issues.

How it lands. This PR is merged first, as is (CI green on 3.10/3.12, Opus review APPROVE, dependents gate in the body). The follow-up branch fix-variadics-all-pk is stacked on it and is landed right after, as its own PR, with the dependents sweep re-run on the combined result.

One note for the record: with *args, excess keyword arguments are merged into **kwargs; without *args, they are ignored (what meshed's hook_up relies on). That asymmetry is pinned by this PR's tests and left as is; it isn't in scope of #78 or #93, and is now tracked in #95.