Fix twelve findings from the 2026-08-17 review

[farrisric]

Fixes the twelve high- and medium-severity findings from the 2026-08-17 review of the v1.4.0 changes (the review ran against PR #21, whose content was already rebased onto main).

Replica exchange

• De Broglie cross-term in the swap criterion. The grand-canonical stationary weight carries Lambda_s^(-3 N_s) per exchangeable species, and Lambda depends on T, so swaps between rungs at different temperatures with different particle counts were biased (e.g. 300/400 K with dN = 10: acceptance off by ~e^4.3). Both BatchedReplicaExchange._accept_swap and the MPI ReplicaExchange._exchange_prob_T now add 3 * sum_s dN_s * ln(Lambda_i,s / Lambda_j,s); lambda_dbs rides on get_state(). Mu ladders were already exact (the term cancels at shared T) and are pinned unchanged; LJ units (Lambda = 1) are unaffected automatically.
• Non-finite swap exponents warn and reject. The old min(1, exp(nan)) form silently accepted every swap involving a NaN energy; the rewritten comparison would have silently rejected. Neither is acceptable for a diverged relaxation, so a NaN/inf delta now logs a warning naming the pair and rejects.
• Per-rung dead-ladder detection. The whole-run degeneracy check could not see a ladder dead at individual rungs (early free swaps inflate the cumulative tally forever). The run now snapshots the tallies at its midpoint and warns per rung on zero second-half acceptance, the test the ladder-spacing doc prescribes.
• Degeneracy warning counts swap attempts, not doubled per-slot tallies (40/40, not 80/80), and is emitted as a single log record so filtering handlers cannot strand the reason line.
• Replica isolation is checked at construction. A factory written for the MPI class (module-level shared atoms/moves, safe under one-process-per-rank) silently corrupts every replica under batching; sharing atoms, move_selector or cells across replicas now raises.

State and cells

• Lossless restart round-trip. set_state restores step and the exchange tallies when the dict carries them; ReplicaExchange strips those slot-bound keys before applying a swap (previously set_state dropped them unconditionally, silently renumbering restarted runs from step 0).
• is_point_exchangeable defaults on BaseCell, and the molecule moves fall back to is_point_inside for plain duck-typed cells, restoring the pre-1.4.0 single-predicate contract that 1.4.0 broke with an AttributeError.

Alchemi calculators

• head= with a pre-loaded MACEWrapper raises. It was silently ignored, so a fine-tuned multihead model shared as a wrapper evaluated on head 0, the pretrain head.
• The energy_only guard message no longer suggests an impossible workaround ("set energy_only on all of the sharers" trips the same guard).
• Forces-stripped models are refused where forces are needed: AlchemiFCalculator.__init__ rejects a wrapper whose active_outputs lacks forces (the .model of an energy-only calculator), and run_md on an energy-only calculator raises up front instead of dying steps deep inside nvalchemi.

Docs

• The ladder-spacing reference implementation now includes the mu_max endpoint (the np.arange fencepost dropped the top rung by up to a full spacing), sorts descending isotherms (desorption scans), and clips noise-negative dN/dmu instead of propagating NaN into a cryptic crash. Verified by executing the snippet: endpoint exact, noisy input stays finite, descending input matches ascending.

Tests

11 new tests: analytic swap probabilities with and without the cross-term, slot-order symmetry, NaN rejection with warning, per-rung second-half detection, un-doubled warning counts, replica-isolation refusal and acceptance, restart round-trip for both ensembles, and the duck-typed-cell fallback. Full suite: 198 passed; flake8 clean.