coef: two-head sigmoid (λ_soft, λ_hard) — .cvcnav format v2

[transfix]

What

Extend the deployable CoefMLP / coef_export / coef_train path with the paper's two learned, sigmoid-bounded reroute strengths — lam_soft and lam_hard (App A.4 / material_nav.py:175-176, lam = lam_max*sigmoid(head)). Previously only lam_soft was learnable (a single softplus 4th output); lam_hard was a fixed dial. This makes the tiny deployable net two-head, byte-compatible with the new libcvc .cvcnav format v2.

Companion (required host): transfix/libcvc — nav: two-head sigmoid (lam_soft, lam_hard) — .cvcnav format v2 → transfix/libcvc#436

Changes

• sdf_nav.CoefMLP: use_lam_hard (requires use_lam) → out_dim = 3 + use_lam + use_lam_hard; lam_soft_max/lam_hard_max = 5.0/10.0. A two-head net is lam_sigmoid: lam columns are lam_max*sigmoid(head) with the head init in the last Linear bias (logit(init/max)); α/β/γ stay softplus(log(expm1)). full_out_bias() returns the 3 abg biases only in sigmoid mode; coeffs_and_lam returns (a,b,g,lam_soft[,lam_hard]). The v1 single-softplus lam path is unchanged.
• add_lam_heads(model, soft_init, hard_init): basin-preserving lifter mirroring add_lam_head — appends the two lam rows (zero weights, bias = logit(init/max)), copies the abg rows verbatim so α/β/γ are bit-identical at init.
• coef_export.write_coef_mlp: a sigmoid net writes FORMAT_VERSION = 2 + FLAG_LAM_SIGMOID, out_bias length 3 (abg only), then the two ceilings (struct '<ff') before the meta trailer. A plain / single-softplus-lam net stays v1, byte-unchanged.
• coef_train: --learned-lam-hard (implies --learned-lam) builds the net via add_lam_heads; --lam-hard-init (>0) seeds the hard head. train_bicycle / eval_risk_exposure use the learned lam_hard for a two-head net; the Swarm drives with both learned lambdas. Deploy note states the 5-out v2 net + the v2 libcvc host floor.

Byte parity / back-compat

• Added a pure-numpy .cvcnav decoder in tests/test_lam_head.py that reproduces the C++ forward from the exported bytes and asserts parity (~1e-4) + lam bounds — an independent cross-check that runs without a pycvc build.
• tests/test_coef_mlp_parity.py gains a pycvc-gated in-process C++↔torch round-trip for the two-head net (all 5 columns, ~1e-4) — runs once pycvc is rebuilt on v2 libcvc.
• v1 single-softplus export stays byte-identical (test_single_softplus_lam_still_exports_v1); all existing lam tests pass.

Test results (.venv, torch 2.12, numpy 2.4)

• tests/test_lam_head.py: 14 passed (7 new two-head).
• Regression sweep (test_swarm, test_scenario, test_squad, test_vehicle, test_material_scenario, test_drive_step_parity, test_sim_thread, test_material_parity, test_vehicle_refinements, test_coef_export_checkpoint, test_coef_train_lam_deploy, test_risk_lever, test_swarm_risk_drive, test_scorecard_eval, test_coef_eval, test_coef_mlp_parity): 130 passed, 9 skipped (all skips are pycvc-only).
• black 24.10.0 + ruff 0.6.9: clean.

NOT in this PR (human owns)

Retrain / A-B / republish (step 9). Ready-to-retrain command:

grl-snam coef-train --rollout bicycle --w-risk 0.5 --learned-lam --learned-lam-hard \
  --lam-soft 0.4 --lam-hard-init 1.0 --steps 800 --seed <S> --out coef_mlp_riskaware_2head_800s<S>.cvcnav

The published libcvc-matext + cvc-dbg-weights recipes will need a cvc_revision bump at republish (not touched here).