Seed H-abstraction TS reactive-core bonds per element

[calvinp0]

Seeds the two reactive-core bonds of an H_Abstraction TS guess from the element that carries each
bond, instead of scaling both by a single factor.

The problem

combine_coordinates_with_redundant_atoms set both the A–H(*2) breaking bond and the H(*2)–B
forming bond to 1.2 × the corresponding equilibrium single-bond length. One factor cannot suit
every element: applied to a heavy X–H it under-stretches the core, and applied to the other end it
is calibrated on nothing in particular.

What replaces it

get_h_abs_reactive_bond_length(symbol, bond) returns get_single_bond_length(symbol, 'H') plus a
per-element elongation — the breaking bond keyed on the donor A, the forming bond on the acceptor B.

The elongations are medians of d − r_e(X–H) over 1670 DFT-optimised H-abstraction transition
states, with r_e taken from the same tabulated source the code adds it back to, so the rule
reproduces each measured median to within 0.003 Å.

element	n (donor)	breaking Δ, 95% CI	n (acceptor)	forming Δ, 95% CI
C	1273	0.223 [0.219, 0.227]	1340	0.282 [0.280, 0.287]
N	207	0.190 [0.181, 0.199]	150	0.259 [0.245, 0.273]
O	138	0.232 [0.221, 0.249]	117	0.301 [0.273, 0.330]
S	30	0.138 [0.124, 0.172]	22	0.233 [0.200, 0.257]
pooled	—	0.218	—	0.281

C, N, O and S carry entries; every other element takes the pooled defaults. Chlorine is not given an
entry — it appears once as an acceptor and never as a donor, which is an anecdote rather than a
median. Hydrogen needs no entry either: the elongation is added to its own r_e(H–H) = 0.75, so the
defaults already yield 0.97 / 1.03 Å.

Subtracting r_e leaves an elongation that is nearly element-free, which is Pauling's bond-order
relation with a fitted order of n ≈ 0.43 (breaking) and n ≈ 0.34 (forming). The textbook n = 0.5
would give +0.18 Å; the data says +0.22 / +0.28. The CH₃ + CH₄ D₃d saddle at 1.34 Å is an
independent check — its elongation of 0.25 Å is the symmetric average of the two.

Conditioning on the donor–acceptor pair rather than on single elements was tested and rejected as
over-fitting: element-only residuals are ≤ 0.05 Å for every pair with n ≥ 80, and the larger
residuals occur only in cells of n ≤ 21.

Accuracy

Mean absolute error against all 1670 reference distances, both bonds:

rule	all records	the 62 H-partner records
1.2 × multiplicative	0.0669	0.1103
this branch	0.0518	0.0908

This is a net improvement, not a uniform one. For the smallest H-transfer systems the multiplicative
factor happened to sit closer to their own saddles — CH₄ + H has a true saddle at 1.383 / 0.904 Å
(verified in the reference set against a CCSD(T) 1.397 / 0.892), where 1.2 × gives 1.311 / 0.897
and this rule gives 1.310 / 1.030. The gain is in the heavy-atom bulk.

The forming elongation exceeds the breaking one because 70.3% of the reference set is written in the
exothermic direction. The asymmetry therefore describes a reaction written in the RMG template
direction and is not a property of the saddle, which is symmetric for a thermoneutral abstraction.

Passing r1_stretch or r2_stretch explicitly still selects the multiplicative path.

Reuse

Searched before writing: get_code_context for existing bond-length and reactive-core seeding,
git grep -n "def get_single_bond_length", and the element data in arc/common.py. No new element
table was added — get_single_bond_length already tabulates r_e(X–H) for every element needed,
and the elongation dictionaries hold only the four measured offsets. No geometry maths was
hand-rolled; the change is confined to the distances handed to set_zmat_bond().

Tests

Four new tests in arc/job/adapters/ts/heuristics_test.py covering a carbon acceptor, an oxygen
acceptor, a hydrogen acceptor and a hydrogen donor, asserting the resulting seed distances.
48 pass in that file, serially and under -n 4 --dist worksteal.

Validated by mutation rather than coverage — five mutants, each killed by a test: flipping an
element branch, reverting to a flat 1.29 / 1.36, reverting to 1.2 ×, dropping the
r1_stretch / r2_stretch escape hatch, and using the donor element for both bonds.

🤖 Generated with Claude Code

https://claude.ai/code/session_01MbGeU8wLpafo3YFzTky2ho

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 65.67%. Comparing base (b7aa476) to head (b8a4107).
⚠️ Report is 3 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #1034   +/-   ##
=======================================
  Coverage   65.67%   65.67%           
=======================================
  Files         121      121           
  Lines       41018    41036   +18     
  Branches    10549    10553    +4     
=======================================
+ Hits        26937    26950   +13     
- Misses      11066    11068    +2     
- Partials     3015     3018    +3     

Flag	Coverage Δ
functionaltests	65.67% <ø> (+<0.01%)	⬆️
unittests	65.67% <ø> (+<0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.