Give the TOA comparison the clock state its own bias needs

[qmohsu]

Closes #85.

run_inline_comparison injects a shared 1.5 m receiver clock bias into the TOA pseudoranges — correctly, to show TDOA differencing it away — then solved them with a position-only (x, y) state. A common bias is unobservable to that state, so the residual never reached tol and the solve was discarded.

Confirmed, and one correction to the issue

Measured convergence, holding everything else fixed:

injected bias	converged	median error at zero noise
1.5 m (before)	2–5 / 100	0.153 m
0 m	99–100 / 100	0.000 m

The bias is the entire cause.

The issue says "the 2–4 % that do converge are accurate". They are not. TOA printed 0.153 m of median error at zero measurement noise, where TDOA and AOA both printed 0.000. A method solving perfect data is exact, so a nonzero number there is never noise. The survivors are the geometries where the bias could be partly absorbed into the position — the least-inaccurate rather than the accurate ones.

That matters for the choice of fix: it rules out option 3 (relabel the panel), because the median column is contaminated too, not just the convergence panel.

Option 1, cheaper than the issue estimates

toa_solve_with_clock_bias is Eqs. (4.24)-(4.26), already in core/rf/positioning.py and already in core.rf's __all__. The example injected a clock bias and then declined to use the solver written for it, one import away.

noise	converged	median error	recovered bias
0.00	100/100	0.000 m	+1.500 m
0.05	100/100	0.044 m	+1.497 m
0.10	100/100	0.091 m	+1.506 m
0.20	100/100	0.158 m	+1.528 m
0.50	100/100	0.464 m	+1.539 m

Summary-table medians go 0.153 → 0.000, 0.177 → 0.044, 0.211 → 0.152. TOA now tracks TDOA at the cost of the extra unknown, which is the comparison Chapter 4 exists to make: TOA carries the clock, TDOA differences it away.

The RSS branch keeps iterative_ls — it reaches the same class through rss_to_distance and carries no clock bias, so it has nothing to estimate. (The issue lists :482 as a position-only site; it is, but correctly so.)

The fix introduced a mirrored defect, also fixed

Three methods then sit on 100 %, and four solid lines at one value show only the last one drawn — the panel that used to say "TOA barely converges" would have said "TOA is missing", which is the same failure this file warns about at :81. Each method now has its own dash pattern. No value is nudged.

Guard behaviour

The chapter README pins this example's stdout, so test_readme_example_output.py went red on the old numbers exactly as designed. The transcript is updated from a real run and passes.

Verification

44 passed across tests/ch4_rf_point_positioning and the lint ratchet; README transcript guard passes; pyflakes clean. Figure regenerated and opened.

🤖 Generated with Claude Code