[thclark]

Closes #11.

Builds the conditioning check into the solver and gives it the validity CLI designed in #11's design comment. Anomaly detection (Class D) follows the same pattern and is #12, to be done separately.

Twins in the solver

• Every analysed point is also solved on five copies of its panels, jittered by −1, 0 or +1 ULP (seeds 1–5, the jitter the reference twins use). The five run in parallel, from the point's own initial state, so its verdict depends only on what its content id hashes. They never steer the base solve.
• A point that is valid on Classes A and B but whose twins spread a reported quantity beyond 1e-4 is Invalid (IllConditioned), and its numbers are withheld.
• Beyond only 1e-8, it stays valid and carries a warning.
• Only values decide. A twin that did not converge raises a TwinsNotValid warning, never a reason.
• The conditioning study (studies/conditioning, docs/validation/conditioning/) sets both thresholds in the gaps of the spread distribution, measured over the series and branch-coverage cases. Over 635 measured points, 12 are withheld and 19 get a warning. Its figures are drawn with ironlab 0.16.

The validity CLI

Flag	What it does
--conditioning on|off|study	on (default) runs the twins. off runs none, as XFOIL. study also writes every twin's record. Replaces --no-twins.
--allow-invalid	Unchanged and orthogonal to every check: shows the numbers of invalid points of any class in results. This is what the issue's REPORT became.
--halt-on CRITERION[=N],...	polar only. Halts a leg after N consecutive points meeting a criterion: xfoil, out-of-domain, unconverged, ill-conditioned or none. Each criterion keeps its own count.
--spread-warning-threshold, --spread-error-threshold	Thresholds for Class C; 1e-8 and 1e-4 by default.

• Study records go in a study array, with role: twin, twin_of, seed and an id of their own. With analyse -o or polar --distributions, each record also carries its twins' full records: jittered geometry, surface and boundary layer.
  • A study's records always carry their numbers, whatever their status, so everything computed can be traced. results is unchanged by a study.
• classes on every record: out-of-domain (A), unconverged (B), ill-conditioned (C).
• The default --halt-on xfoil is XFOIL's NSEQEX rule exactly, so default sweeps don't change. It is narrower than unconverged, which also counts the inviscid Newtons and converged points that passed through a NaN.
  • PolarConfig.max_consecutive_failures becomes halt_on, with a classifier supplied by the caller, because the thresholds belong to the output layer.
  • Halting decides which alphas are attempted, never the numbers of a point that is. The polar records each halt under halts.
• Contradictory flags are usage errors, for example --halt-on ill-conditioned with --conditioning off. --halt-on anomalous is refused with a pointer to Add intelligent polar stepping algorithm #12.
• Ids: base ids are unchanged. Twin ids hash the twin's own panels and seed.

On NACA 4412 at Re 3e6, --halt-on xfoil,ill-conditioned stops the downward leg at −18°, the point the conditioning study flags as ill-conditioned.

Docs

CLAUDE.md, the analyse and polar command pages (new section Halting a leg), solution validity (new sections Classes on records and Halting a polar, plus the study), the data model (brought up to date), terminology and the XFOIL name mapping (NSEQEX → HaltRule::XFOIL).

Test plan

• cargo test --release --workspace: 376 passed, 0 failed. New tests cover twin ids, study records, classes, halt rules and the CLI's usage errors.
• cargo fmt --check and cargo clippy --workspace --all-targets --all-features -- -D warnings
• ci/no-deviations.sh, ci/validity-write-only.sh, ci/ignore-drift.sh
• scripts/docs.sh build --clean and scripts/docs-check.py
• CI