Deprecate positional-only hook implementation parameters

[adenzhou1350]

Refs #774. Revised to follow the agreed deprecation policy: #774 (comment)

Emit DeprecationWarning when constructing a hook implementation with positional-only parameters. Exclude bound self/cls receivers and leave hookspecs unchanged. The warning points to the implementation definition and mentions pluggy 2.0.

The earlier argument-binding fix is removed: _execution.py is identical to current main. Required positional-only calls still work; defaulted positional-only calls retain their existing behavior, including the existing TypeError when supplied by keyword.

Tests cover plain hooks, both wrapper styles, methods, class/static methods, decorated and callable call_extra implementations, and receiver-only controls. The new matrix is 12 failures/1 passing control on unmodified main; all 13 pass with the warning.

Validation: full suite 232 passed on Windows CPython 3.14 and Linux CPython 3.12, using exact source/import checks and existing private dependency environments (pytest 9.0.3/9.1.1), not the complete uv-locked matrix. Mypy and the cached non-Ruff pre-commit checks pass. Local Ruff remains blocked by Windows policy; current hosted checks are pending.

AI-assisted with OpenAI Codex. No performance claim.

[RonnyPfannschmidt]

🤖 Written by Claude Opus 5.5 via Claude Code for the pluggy maintainers; I prompted it, it did the work, I read it.

Approving. I measured the runtime cost against main before approving. The setup was CPython 3.13, a micro-benchmark, 2 rounds, and the best of 15 runs for each case:

Case	main	this PR	Δ
Hook call, 10 impls, no defaulted args	~4.3 µs	~4.35 µs	noise
Hook call, 10 impls with b=None, c=None	~7.2 µs	~7.9 µs	+8–12% (~60–80 ns per impl)
Hook call, 10 impls with positional-only defaults	TypeError	~15.2 µs	new path, ~1.2 µs per impl
HookImpl() construction	~2.55 µs	~4.35 µs	+70% (~1.8 µs per impl)

• Impls without defaulted args are unaffected, because all the new code is inside if hook_impl.kwargnames:.
• Impls with defaulted args now pay on every call for the _posonlydefaults truthiness check. They also pay for a not in _posonlydefaults lookup for each kwarg inside the kwargs comprehension.
• The registration cost is one-time and negligible in absolute terms.

Follow-up (not blocking, will be folded into a refactoring PR stack): move the split to HookImpl.__init__. That means precomputing a kwargnames tuple without the positional-only names for the kwargs comprehension, plus a precomputed (name, default) sequence for the positional-only fill. The call path then only touches the positional-only branch when it is non-empty. With that, impls with ordinary defaults should cost the same as on main, and only positional-only impls pay extra.

---

Generated by Claude Code

[bluetech]

This fix is too complicated and puts the code in the wrong place.

I think it can be done better, but another option is to just disallow positional-only. I'll leave a comment on the issue.

[adenzhou1350]

Thanks. I'll hold off on further changes here until the positional-only policy is settled on #774. If rejection is preferred, I can replace this with registration-time validation rather than extend the call path.

[adenzhou1350]

Updated to the policy agreed on #774: warn for positional-only implementation parameters, keep existing call behavior, and leave the call loop unchanged. Windows/Linux full suites pass 232 tests; the new warning tests fail on unmodified main. The PR description now reflects the narrower scope and validation limits.

[adenzhou1350]

The remaining Read the Docs failure happens before the documentation build starts: build 35160460 fails at git checkout cd600f4 with "reference is not a tree". GitHub refs/pull/775/head currently resolves to the full cd600f4 commit, which I verified with git ls-remote. The test/package/benchmark and pre-commit checks are green. Could a maintainer retry the external docs build? Log: https://app.readthedocs.org/projects/pluggy/builds/35160460/