qFromCharPerActuator: derive Q[3,3] (sigma_do_freq) from char ADEV

[bobvan]

Summary

The foundational fix Main re-sequenced to first: make Q[3,3] (sigma_do_freq) honest by deriving it from the freerun characterization, so the 9.2 ns band-aid can be dropped and √(P22) becomes trustworthy for the OCXO gate, the longTauGnssCoupling coast-cap, and the derived-convergence state.

The bug: derive_do_process_noise derived sigma_do_freq only from an adjfine source — which do_freerun_char.py never writes. So for every real freerun char, Q[3,3] silently fell back to the 0.01 default while only Q[2,2] (phase) was derived. Q[3,3], not Q[2,2], is the coast/convergence knob (the freq-state uncertainty integrates into phase: Var_φ ≈ q_f·τ³/3) — confirmed three ways (Main's analysis, my routedQErrArm tick-discrimination coupling, and Charlie's closedLoopServoSim diverging deterministically on it).

The fix (do_state.py): when there's no adjfine, derive from the rising-ADEV (RWFM) tail:

RWFM Allan variance:  σ_y²(τ) = q·τ/3,  q = Q[3,3] (ppb²/s)
  ⇒  sigma_do_freq_ppb = √3·ADEV(τ)·1e9/√τ

evaluated at the largest available τ (RWFM dominates the long-τ tail; where it doesn't, ADEV(τ) sits above the true RWFM contribution → over-estimates Q[3,3], the safe direction). White-FM↔Q[2,2] / RWFM↔Q[3,3] matches the canonical two-state clock model. adjfine stays primary when present.

Engine: removes the universal 0.92 literal for a per-actuator-structured fallback (behavior-preserving — deliberately not inventing a pessimistic uncharacterized-OCXO constant; that's a sim-validated follow-up). Adds a DOFreqEst Q: summary log showing phase/freq values + sources.

Lab impact (Main coordinates)

• PiFace / clkPoC3 (have freerun chars w/ ADEV): Q[3,3] now derives from ADEV instead of 0.01 on next run — the intended honest-Q change. For a quiet OCXO this is tighter than 0.01, which is correct: an honestly-quiet DO is allowed to coast longer; divergence only happens when Q is dishonestly small vs the true noise.
• MadHat: no freerun char yet → keeps the fallback. Run do_freerun_char on OCXO-33, then test-then-drop the 9.2 ns band-aid (EFC saturation, the root it masked, is already fixed).

Test plan

• 7 new tests (RWFM formula, largest-τ selection, adjfine precedence, non-finite/non-positive skipping, no-ADEV fallthrough, ADEV from a non-chosen phase source, realistic freerun-char schema). Full suite 1387 passed.
• closedLoopServoSim A/B (Charlie): with honest Q[3,3] from ADEV, confirm the >60 s coast no longer diverges — then wire coast_cap_from_p22 (PR longTauGnssCoupling: coast-cap + graded-taper primitives (unit layer) #80) on the now-honest √(P22).
• MadHat: do_freerun_char on OCXO-33 → honest Q → drop 9.2 ns band-aid (Main, lab).

Sequencing: this lands first → coast-cap wiring (PR #80) → sim validation.

🤖 Generated with Claude Code