docs: state the stability policy, and take the free complexity win

[tschm]

Closes #21. Addresses #22 — see the note at the end on how that one should
close.

#21 — a stated stability policy

The surface has been stable since 0.1.x and the README promises a drop-in
replacement, but 0.x carries no semver obligation, so a reader could not tell
whether that stability was a commitment or an accident.

A Stability section now says which it is:

Covered	Not covered
the solve_qp signature — names, order, defaults	anything under _solve / _qr reached directly
Solution field names and order, so 6-way unpacking keeps working	the internal Q/R sign conventions
the two ValueError strings reproduced verbatim from the reference	whether a problem takes the unit-column fast path
the input conventions (linear term subtracted, C column-wise, >=)	bit-exact results, and whether a non-finite G raises or propagates NaNs

Shipping 1.0.0 was the other option the issue offered. That is a release
decision rather than a documentation one, so it is left to you — the section
closes the finding either way, and would need only a wording change if 1.0
follows.

#22 — measured, then took the free half

The complexity was measured rather than argued about, by deleting each
candidate from a scratch copy and re-running radon:

Variant	solve_qp
before	D (23)
extract the argument defaulting	C (19)
…and stub out the inner loop as well	A (4)

That third row is the finding: the inner loop carries 15 of the 23. It is
therefore the only route to a B — and it is also the hot loop, so lifting it out
means threading xv, uv, obj, iact, nact, J, R, u, slack, iter_partial in and
returning five, or hiding them behind an object whose attribute lookups land in
the innermost iteration. The README's own benchmarks show per-iteration dispatch
dominating below n ≈ 160, which is exactly where this package is already
slower than the C reference.

The argument defaulting was the opposite case — separable, runs once per call,
no hot-path cost. It moves to _default_constraints, which takes solve_qp
from D (23) to C (19) and the package average from 4.47 to 4.25.

The "free at runtime" claim was verified, not assumed. There is no
benchmarks/ folder, so make benchmark is a no-op — timed directly instead,
best-of-five on box-constrained problems:

n	before	after
10	0.0236 ms	0.0234 ms
25	0.0290 ms	0.0288 ms
50	0.0631 ms	0.0629 ms

Within noise, in the direction of faster. (Not comparable to the README's table
— different problem setup — but both columns ran the same script minutes apart.)

How #22 should close

solve_qp is C (19), not B, so the issue's first acceptance criterion is not
met and this PR does not auto-close it. Its second criterion is
wontfix with the decision recorded, which is now the case: the comment
above solve_qp carries the measurement, the reason the inner loop is the wrong
thing to extract, and the trigger for revisiting (the function gaining a new
responsibility — a second pivoting rule, an alternative factorisation).

Also

Corrects the layout table, which still described _qr.py as "Givens QR
insert/delete" after insertion moved to a Householder reflection.

Verification

Gate	Result
make fmt	PASS — 21 hooks
make typecheck	PASS — ty + mypy --strict
make test	PASS — 982 passed, 100% statement and branch
make rhiza-test	PASS — 39 passed, 1 skipped
check_test_layout.py	PASS
radon cc src -a	A (4.25), was A (4.47)

The extraction is behaviour-preserving: same 982 tests, still 100% branch
coverage, including the C-without-b and b-without-C error paths that
moved into _default_constraints.

[Copilot]

Pull request overview

This PR documents the project’s pre-1.0 stability commitments for the public solve_qp surface (closing #21), and reduces solve_qp’s measured cyclomatic complexity by extracting argument defaulting into a helper without impacting runtime behavior (partial progress on #22). It also corrects the README’s layout description for _qr.py.

Changes:

• Add a README Stability section that explicitly states what API/behavior is and isn’t covered by compatibility guarantees while 0.x.
• Extract solve_qp’s C/b defaulting and coercion into _default_constraints, reducing solve_qp’s complexity from D (23) to C (19).
• Update the README layout table to reflect Householder insertion + Givens deletion in _qr.py.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
src/cvx/quadprog/_solve.py	Refactors constraint defaulting into _default_constraints and updates the recorded complexity rationale comment.
README.md	Adds a Stability policy section and fixes the _qr.py layout description.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.