Add L2 anchor semantics (initial/uniform/design); fix the flaky refit ESS test

[juaristi22]

Closes #307. Follow-up to #283; experimental evidence in #285 (full results comment incoming).

Why

The l2_lambda penalty divides by a reference vector, and that choice decides which way the penalty pulls: under mass="conserve" its constrained optimum is w ∝ anchor². Anchored at heterogeneous starting weights the penalty concentrates rather than spreads. The post-L0 refit starts from the selection stage's concentrated calibrated weights, so the refit_l2_lambda knob shipped in #283 is a trap at its default anchor: on the full 337k-record US surface a refit penalty of λ=1e-3 collapsed ESS from 8,408 to 1,057 instead of raising it (see #285 sweep data).

What

• calibrate(..., l2_anchor="initial"|"uniform") — "uniform" divides by the mean starting weight, making the penalty a direct 1/ESS control from any starting distribution. l2_anchor_weights is the explicit-vector seam for anchors the frame can't express, recorded in options with a free-form provenance label.
• refit_l0_selection / calibrate_l0_refit gain "design": anchor the refit at the pre-selection initial weights of surviving records ("stay near the survey design") — derived from the selection result, no extra inputs. At a fresh selection "design" collapses to "initial". refit_l2_anchor follows the refit_epochs inherit-or-override pattern; invalid anchors fail before the expensive selection stage.
• Direction contract pinned where it's decidable: from heterogeneous starts (ESS 29.3 of 160), the initial anchor drives ESS below 0.6× the start while the uniform anchor drives it above 1.4×, at healthy loss — margins orders of magnitude above run-to-run noise.

The #307 fix

test_refit_l0_selection_threads_l2_lambda_to_refit asserted an ESS direction between penalized and free refits on a near-converged fixture — a regime where the mass-conservation projection cancels the penalty's near-proportional push, leaving the comparison at the ~1e-5 nondeterminism noise floor (and the direction premise itself inverts at production scale, per the above). Per the issue's own guidance, no padded tolerance: the threading contract is now pinned via recorded solver options — exactly what the original hardcoded l2_lambda=0.0 bug would have violated — and the behavioral direction lives in the anchor test, in a regime where it's real. Verified stable across repeated runs.

Tests

Full suite: 1,343 passed / 47 skipped on current main; ruff check clean. New: anchor direction contract, explicit-anchor exact-equivalence with presets, validation (labels, shapes, positivity, fail-before-selection), two-stage inherit/override provenance including "design".

🤖 Generated with Claude Code