Add the ability to set 'inherit' at all levels of the options cascade

[joerick]

Adds the ability for options set in

• [tool.cibuildwheel]
• [tool.cibuildwheel.platform]
• env vars

to inherit the values previously set.

Previously one could only inherit in an 'overrides' config.

I always wanted this to be possible for env vars, but I also realised that when we were doing audit-command, it would be useful for projects to be able to extend the default audit (e.g. add twine check to the existing abi3audit command). So I think it makes sense to allow this at every user-configurable level.

Docs + tests still to come.

[joerick]

This should be ready for review now!

[henryiii]

Here's a Claude review.

🤖 AI text below 🤖

cibuildwheel/options.py
  ● 568 [correctness]       Moving overrides parsing from __init__ to a functools.cached_property makes override validation lazy, so malformed overrides config is silently accepted on paths that never compute per-identifier options.
  ● 445 [correctness]       parse_inherit validates rule values but never the keys against known option names, so typos in the new global/platform/CIBW_INHERIT inherit surfaces are silent no-ops that degrade to replace semantics.
  ● 405 [correctness]       _apply_inherit_rule does not catch OptionFormat.NotSupported, so an append/prepend inherit rule on an allow_merge=False option (container-engine, xbuild-files, test-runtime — all listed in the new schema's inherit keys) crashes with a blank-message traceback.
  ● 874 [correctness]       An inherit append/prepend rule on xbuild-tools merges the user value with the "" unset-sentinel default, defeating the sentinel check and leaking a NUL tool name.
  ● 599 [correctness]       parse_inherit's string path lets shlex's bare ValueError escape: env_inherit catches only OptionsReaderError, and the config-file/override call sites have no wrapper at all.
  ● 460 [correctness]       parse_inherit's value check `v in {"none", "append", "prepend"}` raises TypeError on unhashable TOML values instead of the intended OptionsReaderError.
  ● 604 [api-design]        Platform scoping for env inherit rules is encoded as a '-<platform>' suffix inside CIBW_INHERIT keys instead of the CIBW_<OPTION>_<PLATFORM> convention every other option uses, and the suffix matches against no vocabulary.
cibuildwheel/util/helpers.py
  ● 189 [correctness]       In CIBW_INHERIT, a forgotten colon or a space inside a rule word is silently accepted with the wrong meaning: bare words become keys with default 'append', and parse_inherit's "".join glues multi-token values back into valid rules.
  ● 150 [simplification]    parse_arbitrary_key_value_string is a near-verbatim copy of parse_key_value_string (16 of 27 body lines byte-identical), added in the same PR that deletes the '*' wildcard which already provided arbitrary-key parsing.
bin/generate_schema.py
  ●  43 [schema-divergence] The schema's inherit declaration diverges from runtime acceptance: string form rejected, the 18-key list omits now-reachable list options (enable, archs, test-groups, build, skip), and platform sections $ref the full inherit object including keys invalid for that platform.

Review of PR #2928 is done — 10 findings reported above, ranked by severity. Eight are confirmed with executed reproductions; two (the schema/runtime divergence and the CIBW_INHERIT platform-suffix design concern) are plausible but not mechanically reproduced.

The broad themes:

• Error handling around the new inherit surfaces is the weak spot. Typo'd keys, missing colons, unbalanced quotes, non-string TOML values, and append/prepend on non-mergeable options all either silently do the wrong thing or escape as raw tracebacks instead of the usual friendly fatal errors. One fix location (parse_inherit) covers several of these.
• A real regression: moving overrides parsing into a cached_property makes validation lazy, so --print-build-identifiers now exits 0 on config that previously failed fast.
• Two pre-existing bugs got widened: the OptionFormat.NotSupported crash and the \x00 xbuild-tools sentinel leak existed via override inherit, but the new top-level/platform/env surfaces (and the schema advertising the keys) make them much more reachable.
• Cleanup: the new parse_arbitrary_key_value_string duplicates the existing parser almost verbatim and could be a default_value kwarg instead; and the schema's 18-key inherit list no longer matches what runtime actually accepts.

The reviewer also verified the checked-in schema matches bin/generate_schema.py output and that all 220 unit tests pass at PR head. Since the CIBW_INHERIT env var is brand new in this PR, the naming/validation concerns (findings about key typos and the -<platform> suffix) are cheap to address now and costly after release — worth deciding before merge.