P0 physics claim-boundary debt: T2b demotion, fitted-state sweep, clamp audit

[neuron7xLab]

Goal

Close newly identified physics-governance and model-semantics debt without granting score credit from documentation text.

Findings

P0-1: INV-T2b claim-boundary drift

results/cross_asset_kuramoto/SL_ES_PREREGISTRATION.md records DECISION: REJECT H1 for the Stuart-Landau ES lead claim:

• leads_rate = 0.90625
• p_value = 0.7647647...
• n_episodes = 32
• rejection reason: circular-shift null makes the apparent lead indistinguishable from chance

.claude/commit_acceptors/T2b-stuart-landau-es.yaml correctly says no T2b signal is wired into the shadow rail.

But .claude/physics/INVARIANTS.yaml still exposes INV-T2b as a qualitative P1 invariant-like physics claim. That creates claim-boundary drift: a rejected empirical hypothesis remains in the hard physics registry.

Required fix:

• demote/mark INV-T2b as REJECTED_DO_NOT_PROMOTE or move it out of invariant registry into a rejected-hypothesis ledger;
• preserve INV-SL1 and INV-SL2 as universal implementation bounds;
• add a guard preventing rejected empirical hypotheses from appearing as active invariant claims.

P0-2: Stuart-Landau sweep ignores fitted market state

fit_stuart_landau() extracts fitted market amplitude, phase, omega, and mu, but _hysteresis_sweep() receives only mu and omega; it initializes z from seeded random normal values.

That means ES proximity can become a function of mu/omega + seed, not a full sweep of the observed market analytic state.

Required fix:

• pass fitted z0 = amplitude * exp(1j * phase) into _hysteresis_sweep();
• keep seed only as explicit fallback/noise mode, never as the default fitted-state path;
• add a regression test showing two windows with same mu/omega but different amplitude/phase produce distinguishable sweep state when physics says amplitude dynamics matter.

P0-3: silent μ clamp without audit surface

_estimate_growth_rate() computes μ as a rate, then clamps to ±1 for Euler stability. That may be a necessary bounds guard, but it silently changes the physical growth-rate distribution.

Required fix:

• expose mu_raw, mu_clamped, mu_was_clipped, and clipping count/mass in result or diagnostic artifact;
• add a test that pathological μ is clipped visibly, not silently;
• forbid interpreting clipped windows as unmodified physical evidence.

P1-1: rolling physics failures become NaN

rolling_es_proximity() catches ValueError / FloatingPointError and writes NaN. For exploratory plotting this is acceptable; for physics evidence it hides failure reasons.

Required fix:

• add strict mode: fail_closed=True raises with window index and reason;
• add optional diagnostics ledger for NaN windows;
• evidence-generation path must use strict mode.

Non-claims

• Do not raise physics score from this issue alone.
• Do not promote T2b to shadow rail.
• Do not weaken existing gates.
• Do not erase the rejected OOS result.

Acceptance criteria

• Active registry cannot present rejected empirical hypotheses as live physics invariants.
• Stuart-Landau ES proximity default path is initialized from fitted analytic state, not random seed.
• μ clipping is auditable.
• Rolling failures are fail-closed in evidence mode.
• Tests and artifacts distinguish implementation bounds (INV-SL1, INV-SL2) from rejected lead claim (INV-T2b).

[neuron7xLab]

Runtime execution scope for #1358:

• Use branch t2b-runtime-fix.
• Make Stuart-Landau sweep use fitted analytic state from amplitude and phase by default, not seeded random initialization.
• Add visible mu clamp diagnostics: raw, clamped, clipped flag/count/mass.
• Add strict rolling mode that raises with window index and cause instead of silently writing NaN.
• Add tests for seed-invariant fitted sweep, visible mu clipping, and strict rolling failure.
• Do not change score, trading, execution, forecast, policy, or shadow-rail status.

Primary check: python -m pytest tests/unit/physics/test_T2b_stuart_landau_es.py -q.