[wshlavacek]

Closes #1063.

#1109 and #1188 moved rm runs and sessions of a model that reads time() through its functions onto elapsed time. They left one case, a model with a table function indexed by time, on model time: offsetting a table's clock would index it through a function, which RuleMonkey 3.11.0 could not evaluate (richardposner/RuleMonkey#111). On model time, #1022's require_resolvable_clock refused any run, segment or step whose span the doubles near its times could not divide into RuleMonkey's integration panels, for example a span of 2 at t = 1e13.

RuleMonkey 3.12.0, vendored in #1220, keeps its clock as a per-call origin plus elapsed time. It places events, samples and its panels for a rate that reads the clock on the elapsed part (richardposner/RuleMonkey#110), and rounds only the clock a rate reads to the doubles near it. The table case needs no offset any more, and the refusal now turns away runs RuleMonkey gets right. This PR removes it from run(), step_to() and simulate().

Measured

The fixture is bngsim's step table, which reads 20 for t ≥ 2, seeded with 10⁴ X(p~1) and dephos = 0.001, for a total propensity of 2×10⁵. The table shows X_phos at t0 + 0.05 over 20 seeds, against the closed form on the span the doubles near t0 hold:

t0	main	RuleMonkey 3.12.0 alone (rm_driver --t-start)	this branch
3	z = −1.1	z = −1.1	z = −1.1
1e9	refused	z = −1.1	runs
1e12	refused	z = −1.1	runs
1e13	refused	z = −1.2	runs

Tests

• New: test_a_time_indexed_table_keeps_its_waiting_times_far_from_t0 runs that model one-shot and as a RuleMonkeySession from 1e12 and 1e13, 40 seeds each, against the closed form. All four cases fail on main with the refusal.
• Converted: test_a_clock_too_coarse_to_integrate_over_is_refused pinned the refusal, and becomes test_a_short_span_far_from_t0_runs. The table model now runs that span one-shot, as a session segment and as a step. The test's half for clock-reading functions is unchanged.
• RuleMonkey tests (the 18 files that drive it): 500 passed, 3 skipped. That is 496 before plus the 4 new cases.
• Full suite: the pre-push hook's full python/tests run passed.

Corrections to #1220 and #1063's fragments

• +rulemonkey-3-12-0.fixed.md credited the re-vendor with "a model that reads the clock keeps its waiting times from a large t_start". The model I measured that on read time() through a function, which rm: run a model that reads time() on elapsed time, with time() offset by t_start (#1063) #1109 had already moved onto elapsed time. The case 3.12.0 changes is the one here. Its sentence about Arrhenius rules staying refused until Lift method='rm''s refusal of Arrhenius rules RuleMonkey misreads, once richardposner/RuleMonkey#109 is fixed and vendored #1058 is dropped too, since rm: run the Arrhenius rules RuleMonkey 3.11.0 misread (#1058) #1221 lifted them.
• 1063.fixed.md said a table "still runs on model time" because of RuleMonkey No first-class way for an on_point dose override to declare a θ-dependent initial condition, so its seed row is silently zeroed #111. It now says what 3.12.0 does.

docs/user-guide/network-free.md, the set_reads_clock header comment and binding docstring (with the regenerated .pyi), and the comments in _rulemonkey_clock.py, _simulator.py and rulemonkey_simulator.cpp say the same.

Not changed

#1059 stays open. method="ssa" keeps its rounding limit on a rate that reads time(), which runs on the continuous loop (#1075's "Not covered"). An SSA event whose trigger reads the clock can also still fire up to half a spacing early.