feral v0.6.0

Added — Solver::with_scaling() builder ([#51][i51])

New public builder method on Solver to pin the scaling strategy
explicitly. Lets IPM hosts override ScalingStrategy::Auto when
they already know what they want — recommended escape hatch when
the picker's per-matrix heuristic is wrong for a problem class.
Default (Auto) is now sticky on cached pattern (see "Fixed" below)
so the picker no longer flaps; this builder is for callers that
want full control.

Fixed — sticky Auto scaling on cached pattern ([#51][i51])

Three coordinated fixes in src/numeric/solver.rs address a ~50×
numeric-phase slowdown on IPM workloads when the picker re-routes
across iters or PartialSingular outcomes drop the MC64 cache.

Reproducer (pounce/gams/nlpbench/feral_repro/powerflow22/,
n=2,813,976 IPM KKT, default Auto):

call	pre-fix	post-fix
factor (cold)	54.77 s	54.42 s
refactor (cached symbolic, iter A values)	1.00 s	1.02 s
factor #2 (cached symbolic, iter B values)	53.80 s	1.07 s

The three fixes:

1. Sticky Auto pick. First factor() on a pattern runs the full
   compute_scaling_auto_with_cache pipeline (preserves Policy-4
   fallback semantics — InfNorm-spread guard, off-diag-ratio guard,
   MC64 catastrophic-spread guard). Post-call we derive the resolved
   strategy from factors.scaling_info
   (Mc64FallbackToInfnorm → InfNorm,
   PartialSingular → Mc64Symmetric, NotApplied → Identity,
   Applied → pick_scaling_strategy(matrix)) and stash it on the
   Solver as auto_picked_strategy. Every subsequent factor on
   the same pattern uses the stashed strategy directly, bypassing
   Auto. Pattern change clears it alongside the MC64 cache. Mirrors
   MUMPS ICNTL(7) / SSIDS options%ordering: structural decision
   once at first call, reuse every refactor.

2. MC64 cache gate widened to PartialSingular. The Hungarian
   on a structurally rank-deficient KKT still produced a real
   scaling vector (unmatched positions land at 1.0 per
   mc64.rs:222); the value-bound check still gates reuse, so
   caching is correctness-safe. Pre-fix the post-#49 gate required
   ScalingInfo::Applied, which dropped the cache on every IPM iter
   over a structurally rank-deficient KKT and forced the Hungarian
   to rerun from scratch.

3. Solver::with_scaling() builder (see "Added" above).

Tests

Three unit tests in src/numeric/solver.rs:

• issue_51_with_scaling_builder_overrides_default
• issue_51_auto_pick_is_sticky_on_cached_pattern
• issue_51_partial_singular_populates_cache

And one #[ignore]'d corpus regression test:
issue_51_corpus_sticky_auto_holds_across_ipm_iters walks
tests/data/parity/<family>/*.mtx for every family with ≥ 2 IPM
iter snapshots, factors all iters against one Solver::new(), and
asserts the sticky pick holds across every iter (13 families
covered: acopp30, hatfldbne, hahn1, ssi, …).

mc64_fallback_surfaces_via_solver_api was updated to reflect the
new sticky semantics: iter 2 now asserts
mc64_fallback_count == 1 (sticky pin runs straight InfNorm; no
fallback to surface) and locks
auto_picked_strategy = Some(InfNorm).

Changed — Auto dispatcher rewrites ([#50][i50])

The OrderingMethod::Auto dispatcher (src/symbolic/mod.rs:: choose_adaptive) was simplified to two rules on top of
pick_default_method: kept only the very-large-and-sparse catch
(n > 100_000 && full_avg_deg < 5.0 → Amd); everything else
delegates to pick_default_method (n ≤ 10_000 → Amf,
n > 10_000 → MetisND).

• Fix A — large-and-sparse swap (c442a0c). The pre-fix
  n > 100_000 && full_avg_deg < 5.0 → ScotchND branch is swapped
  to Amd. On powerflow22 (n=2.8 M, full_avg_deg ≈ 3.7) prior
  ScotchND took 113.8 s symbolic (15.8 M nnz_L); AMD takes 55 s
  (10.4 M nnz_L). The ScotchND advantage at very large n had been
  load-bearing against the same BK pivoting cascade that motivated
  the chain catches; issue #46 eliminated that amplifier in May
  2026 and removed the justification for routing very-large sparse
  matrices through nested dissection. Corpus inventory:
  dev/research/issue-50-numeric-inventory.csv shows the IPM
  corpus's [100k, 200k) bucket has AMD/MetisND num_nnz_l ratio 1.00
  on both representatives. Validation: 258 chain-catch corpus rows
  under post-Fix Auto — 0 failures, 0 num_nnz_l regressions for
  matrices that actually reroute. See
  dev/research/issue-50-metisnd-symbolic-cost.md §F7–§F11.

• F11 follow-up — small-and-sparse retire (3f8f6f6). The
  pre-fix n < 10_000 && full_avg_deg < 15.0 → KahipND branch is
  deleted entirely; the population now falls through to
  pick_default_method's n ≤ 10_000 → Amf default (MUMPS
  ana_set_ordering.F SYM=2 N≤10000). Justified by an 838-matrix
  4-way inventory (dev/research/small-sparse-inventory.csv): AMF
  wins 169/838 strict per-matrix (KahipND 16); aggregate num_nnz_l
  ratios to AMD are AMF 0.870×, KahipND 0.984×; aggregate factor_us
  ratios are AMF 0.832×, KahipND 0.990×. KahipND remains reachable
  via OrderingMethod::KahipND for the 41 high-avg-deg cases
  (STEENBRD, HADAMARD, TABLE8) where it still wins — all sub-22k
  nnz_L absolute. See dev/research/issue-50-metisnd-symbolic-cost.md
  §F12.