[wshlavacek]

Closes #1009. Closes #910.

Follow-up to #1004. This fixes the four findings from that PR's independent review. #1004 was merged at its reviewed head before these were done.

1. Only refuse an output key that something reads (#1004 introduced the false refusal)

#1004 refused any configuration where two (action, condition) pairs spell the same output key. That included a key nothing reads: experiments Egf (condition Inh) and EgfTnf (condition TnfInh) give two unmeasured pairs that both spell EgfTnfInh, and emit pruning runs neither of them.

In edition 2, a colliding key is now refused only when something reads it:

• data keys;
• each experiment's own output;
• constraint keys and cross-references;
• postprocessing targets;
• the keys of actions no experiment created, which always run.

The check runs again after postprocessing targets load. Edition 1 keeps the blanket refusal, since there everything runs.

Tests confirm both directions:

• A read key that collides really is overwritten. With the check disabled, data comes back at k1 = 0.9 instead of 0.3.
• An unread key is never scored.

An unused condition that completes a scored key now gets a hint: every experiment runs under every condition of its model unless nothing reads the result.

2. A pre-equilibration's hidden phase output overwrote a real experiment (older; silent)

A pre-equilibrated experiment x also writes x_preequil. An experiment really named x_preequil had its output replaced. With steady-state data this was silent: its term was 987.654, the phase's steady state, where it should have been about 0. The collision check now counts every output an action writes, registered or not, so that configuration is refused in either order, naming the phase.

3. Unanchored legacy mutant-file match (older; silent)

config.py's re.match(f'.*(?={name})', ename) let xbc.exp under mutant b, with actions x and xbc, load and be scored against the unmutated action xbc. The file stem is now resolved from the model's (action, mutant) pairs and must belong to this mutant; otherwise the configuration is refused. xb.exp and xbcb.exp still load.

4. Duplicate action suffixes

Duplicate suffixes are now refused at load. This covers a legacy time_course key next to experiment: e1, simulate plus parameter_scan on one suffix, and a time_course reusing a model-file suffix. Two kinds stay valid:

• a continue=>1 that resumes the run before it;
• two actions of the same kind in the model file, where BNG2.pl keeps the later output (Miller2025 MEK1_KO.bngl does this).

output_pairs now keeps the last action too, so multiple shooting agrees with execute.

Tests and sweeps

• New tests: 18 in tests/test_output_key_resolution.py, plus the fix(gradient): resolve each output from the (action, condition) pair that produced it; a no-column route requests none (#850, #1002, #833) #1004 reviewer's tests/test_output_keys_review_1004.py. Its formerly-xfail test now passes.

• Before this change: 8 of the new tests and the reviewer's xfail fail on dd4c6a3a.

• Both bngsim builds: the config, edition, emit-pruning, pre-equilibration, steady-state, shooting, gradient-oracle, PEtab and am test files all pass.

• Sweeps: four load sweeps match main exactly:

  Sweep	Configs	Load
  Examples	553	549
  BNGL-Models jobs	102	78
  PEtab imports	24	19
  PEtab benchmarks	5	5

  The failures are the same configs failing with the same messages as on main.

ruff and tools/changelog.py check are clean. ADR-0153 has a new item 20.

After independent review (a5c1ea7, 70efedc)

The first review found a silent regression and a false refusal. Both are fixed, along with three older items the PR text had claimed handled:

• Constraint reads. The "reads" set missed a constraint's enforcement variables (at X.obs=v, between, split-at) and constraints from other models. So a colliding key that only such a constraint reads loaded, and the constraint scored the wrong pair: 1.30079, where 0.2611 is right; main refused this. _constraint_ref_suffixes now reads all nine attributes, and both the reads set and the emit set collect references from every constraint set. The same change fixes an older crash: a separable bngsim model pruned an at-referenced pair.
• Suffix-less warm-up. An unscored suffix-less warm-up line next to a time_course key was falsely refused. It now counts as a second writer only where bngsim may run the model and the line comes after the last registered writer.
• On the bngsim backend, a simulate() with continue=>1 that reuses the previous suffix replaces the earlier rows instead of appending to them, so data on the first segment fail every evaluation (or, with ind_var_rounding = 1, are scored against the wrong rows); a continuation whose t_start differs from the clock runs where BNG2.pl refuses it #1009. On bngsim .net, a continue=>1 sharing a suffix now appends its rows and forward sensitivities, as BNG2.pl does. The reviewer's probe went from 6159 to 2.9e-10 (BNG2.pl 5.3e-11). The gradient through a 3-run chain matches finite differences. A shared-suffix continue on network-free is refused.
• job_type = ms on a .net model integrates only the scored simulate() line and skips the rest of the action protocol (pre-equilibration, setParameter/setConcentration, carried state), so it fits and certifies the wrong experiment #910. Multiple shooting on .net now refuses a scored run that an earlier action feeds with no reset in between: a simulate, a scan, setParameter, setConcentration, or a restored snapshot. That is job_type = ms on a .net model integrates only the scored simulate() line and skips the rest of the action protocol (pre-equilibration, setParameter/setConcentration, carried state), so it fits and certifies the wrong experiment #910's case.
• Docs. Unread pairs may run (on BNG2.pl, RoadRunner, under check, or on a non-separable bngsim model), but nothing reads them.

The second review found two more problems, fixed in 70efedc:

• A continue=>1 with an explicit t_start later than the model time was appended time-shifted (B(7) came back as the closed-form B(5)). It is now refused, as BNG2.pl refuses it.
• Multiple shooting falsely refused an unlabelled resetParameters().

The reviewer's tests are committed. This PR conflicts with #1005 in net_model.py, and #1005's new data-on-the-grid check must learn about appended continue chains. #1005 lands first; this branch will then merge main and resolve both.