Document mir_history offset; fix waypoint assertion

[neuromechanist]

Closes #161, and repairs a test my own #241 broke.

#161 was mostly already done

Both of its testable claims already have tests in test_ng_convergence.py — test_mir_history_survives_keep_best_restore and test_mir_history_empty_after_save_load. Only the third, informational item was outstanding.

Verified before documenting it: ll comes from the E-step accumulator at the top of the loop, _update_parameters runs, and the MIR waypoint is computed after. So mir_history_[i] reflects the parameters after iteration i's update while ll_history[i] is the likelihood of the parameters before it. Zipping them by index compares different parameter states. Now stated on both the backend attribute and the AMICA.mir_history_ docstring.

A regression from #241

Running the whole file rather than a subset surfaced test_mir_history_survives_keep_best_restore failing on dev:

assert 50 == (55 - 1)

I bisected it: the test passes at 257d20e (the commit before #241 merged) and fails at c80fb33 (the merge). #241 caused it. That PR's gm-ordering change shifts multi-model trajectories by ~1e-6, which moved the min_dll stop to iteration 55; MIR waypoints land on multiples of mir_step=5, so the last one is at 50, not 54.

The assertion last_it == len(ll_history) - 1 was fragile from the start — it only held while the stop happened to coincide with a waypoint, and any trajectory perturbation could break it. It now asserts what the test is actually about: the last waypoint is the highest multiple of mir_step at or below the final iteration, and is within one interval of the end, so a restore that truncated the trajectory would still fail it.

Why CI did not catch it

The test is not slow-marked, so CI runs it, and CI was green on #241's branch. I have not established why the branch and the squashed merge behave differently here; the practical lesson is that I ran a subset locally on #241 and would not have seen this either way. Flagging rather than leaving it implicit.

[neuromechanist]

Review findings addressed; merging.

Finding 1 (Critical) — the assertion could not witness a truncating restore. Confirmed and fixed. The reviewer's sabotage (filter mir_history_ to t[0] <= best_it right after _restore_params) passed under the old assertion because min_dll/maxincs halts 1-2 iterations past the peak while mir_step=5, so the last waypoint (50) sat before both best_it (53) and final_it (54) — the damaged window was never sampled. Fixed by making mir_step=1 load-bearing: every iteration is a waypoint, so one always lands strictly inside (best_it, final_it]. Re-ran the same sabotage against the new test: it now fails, and passes again after revert.

Finding 2 (Important) — assert last_it > final_it - 5 was a tautology. Agreed, the algebra is exact. Removed rather than reworded, and replaced with a count assertion (len(mir_history_) == final_it + 1) that catches partial truncation the last-entry check would miss.

Finding 3 (Important) — magic 5 duplicated three times. Dissolved by the fix: with mir_step=1 the assertions are last_it == final_it and len(...) == final_it + 1, so no literal remains.

Finding 4 — noted; the new form is robust to stopping-iteration drift, since it no longer depends on where the stop lands relative to a waypoint boundary.

The docstring accuracy confirmation was useful independently: the off-by-one being documented is real (ll_history[i] pre-update, mir_history_[i] post-update).

22/22 in the file pass, ty clean.