docs: add CLAUDE.md with engineering guidance for AI-assisted changes

[hozblok]

Adds a CLAUDE.md at the repo root documenting the engineering principles AI-assisted changes should follow.

Lead principle. Prefer obvious failures over silent acceptance of wrong-shaped input. If a caller passes the wrong type or shape to a public API, the wrapper must raise a clear, named exception whose message identifies both the parameter and the offending value's type — no coercion, no silent fallback, no sentinel return.

Other sections:

• Right exception type (TypeError vs ValueError vs RuntimeError; no bare Exception).
• Breaking-change contract for public-API tightening.
• Test discipline — every behavior change comes with a regression test that fails pre-change and passes post-change.
• Don'ts — no drive-by reformatting, no touching unrelated files, no loosening type contracts for ergonomics, no bypassing safety checks to clear an obstacle.
• Practical pointers for building the C++ extension and running the suite.

Incoming PR #16 (fix/02-solver-rejects-scalar-no-vars) is being amended to follow the lead principle (replace silent scalar-coercion with a clean TypeError); this doc should land first so that PR's body can link to it.

[hozblok]

Closing as obsolete.

A different CLAUDE.md was added to master independently with substantive content of its own (minimalism / zen style, rules around comment brevity, code simplicity, no flattery, verify-before-assert). This branch's AI-engineering-principles content does not apply as-written on top of what's there now — they're different documents with different framing.

If any of the principles here (loud failures over silent acceptance, exception-type discipline, test-discipline rule, the Don'ts list) are still wanted, they belong as a small additive edit to the current CLAUDE.md in a fresh PR — not as a replacement of it.