benchmarks: settle the argv contract for the whole suites/ tree (#489)

[wshlavacek]

Closes#489. Follows #490, which fixed
the two runners run_all.py can reach.

The paper side's comment on the issue is followed exactly: _ssa_cell.py and
bngsim_wrapper.py get no parser, that exemption is now endorsed by a test
rather than left to the next implementer, and the freedom granted over
recover_s4_points.py is used.

The hazard was demonstrated, not argued

Two of the seven wrote committed files on a bare invocation. Reproduced
against this repository, then restored from git:

$ cd benchmarks/suites/jacobian && python probe_attach.py --help
--help: attached=None n_func_rxns=None ns=None self_check_failed=None err=missing
WROTE results/attach_probe.json
$ git diff --stat -- results/attach_probe.json
.../jacobian/results/attach_probe.json | 48 +---------------------
1 file changed, 2 insertions(+), 46 deletions(-)

46 lines of measurement replaced by [{"model": "--help", "error": "missing"}] —
by the command typed to find out what the script does. That file is the record of
which models genuinely attach the analytical Jacobian, which is what decides
membership in the suite's speedup table.

script	what --help used to do
jacobian/probe_attach.py	probed a model named --help; overwrote the committedresults/attach_probe.json
ode_engines_s4_sbml/check_sbml_engine_agreement.py	ran bngsim + RoadRunner + COPASI over the Table S4 models, then overwrote its committed report
ode_fullnet/recover_s4_points.py	regenerated two networks through BNG2.pl, rewrote the characterization under parity_checks/, then copied both into another repository'slatex/generated/
ssa_table5/convert_all.py	reconverted all 14 models, rewrote results/converted/ + conversion_log.json
ssa_table5/emit_ssa_table.py	rewrote ssa_timing_ballpark.json in place and re-rendered the .md
ssa_table5/merge_jobout.py	rebuilt ssa_timing_ballpark.json from _jobout/
jacobian/diagnose_divergence.py	read --help as a model id and stopped — the one benign case

The fix

All seven build an ArgumentParser: --help exits 0 with usage, an unrecognized
flag exits 2, a bare invocation is unchanged.

Where a probe needs somewhere harmless to write, it now has a flag — the same
shape --out took on the PSA companion in #490:

• --out on probe_attach.py and check_sbml_engine_agreement.py;
• --paper-dir / --no-copy on recover_s4_points.py. That one also skips
  rather than crashes when the paper checkout is absent: the merge has already
  happened by then, so a FileNotFoundError at the copy left the work done and
  the run marked failed. Its stale /Users/wish/Code/PyBNF/... follow-up line is
  gone too, per your note that it was never a contract.

diagnose_divergence.py is both driver and worker — it re-enters itself as
--worker on an argv it composes in _spawn(). That branch is matched before
the parser and left byte-identical; only the half a person types is parsed.

The exemption, endorsed

ssa_table5/_ssa_cell.py (spawned per (engine, model) by run_ssa_timing.py,
Table 5's measurement path) and sbml_test_suite/testrunner/bngsim_wrapper.py
(invoked by the vendored runner on %d %n %o %l %v) get no parser. Each now
names its caller in a comment saying why there is no parser.

Worth noting: bngsim_wrapper.py was already correct. Its argc check answers
--help with a usage line and exit 2 — the issue's table overstated it as
ignoring argv. It needed the comment, not a change.

The endorsement is a test, not a convention:

deftest_spawned_workers_keep_their_positional_contract(rel):
src= (SUITES_DIR/rel).read_text()
assert"argparse"notinsrc, f"{rel} grew a parser -- it is spawned by ..."assertSPAWNED_WORKERS[rel] insrc, f"{rel} does not name its caller"

So the next argv sweep cannot quietly "fix" these two, and cannot read the
missing parser as an oversight either.

The test now discovers its family

test_benchmark_runner_help.py no longer derives from run_all.py's registry
plus a hand-kept companion list. It globs every suites/ script with a
__main__ guard and subtracts SPAWNED_WORKERS — 59 scripts, 127 cases,
15 s. A script added to any suite inherits the contract without an edit here.
A separate test asserts every registry-named script lands inside that family, so
the orchestrator link stays explicit.

Two tests assert --help leaves each committed artifact byte-identical, which
states the actual harm rather than only "exit 0".

Verification

• 127 cases pass; full local suite green.
• Re-ran every script's --help with libsbml, scipy, roadrunner, amici,
  COPASI, antimony, gillespy2, diffrax, jax, pandas, matplotlib and
  psutil blocked at import. Only two scripts import an optional engine at module
  scope — biomodels/filter.py (libsbml) and showcase/run_ode_trf_fit_from_net.py
  (scipy) — and both deps are guaranteed present wherever this test runs
  (python-libsbml is a core dependency; scipy comes in via the test extra,
  which dev includes and CI syncs).
• diagnose_divergence.py --worker verified to still reach worker() and fail
  inside it on a bogus path, proving the spawn argv unpack is intact.
• The two committed artifacts are byte-identical in this branch.

Risk

None to any measured value. Every script behaves identically when invoked as it
was before, and the two scripts on a measurement path are the two that were not
touched functionally.

[wshlavacek]

CI was red — 19 cases, all in the new test file, all one defect it had just been widened enough to see. Fixed in c031852.

Eight scripts resolved their repo root as Path.home() / "Code" / "bngsim" and then put <root>/parity_checks on sys.path. Their module-scope import _bng_common therefore resolved only on a machine whose clone sits at that path, and raised ModuleNotFoundError in any other checkout — a worktree, a fresh clone, every CI runner:

ode_engines_s3/run_s3_timing.py (+ patch_amici_split.py, which imports it)
ode_engines_s4_sbml/run_s4_timing.py
ode_engines_s4_sbml/check_sbml_engine_agreement.py
ode_fullnet/{run_timing,run_forced,gen_networks,probe_growth,recover_s4_points}.py

The default is now Path(__file__).resolve().parents[3] — the checkout the file lives in — which is the idiom biomodels/emit.py already uses. BNGSIM_ROOT still overrides, so the case the env var was written for (a different venv run against a canonical checkout) is untouched, and where the two already coincided the resolved path is identical.

The test could not have caught it, because the author's $HOME made the wrong default look right. Every probe now runs with HOME/USERPROFILE pointed at an empty session-scoped directory. Stashing one script's fix reproduces the exact CI failure locally, so this class is now caught on a dev machine rather than on a runner.

Worth noting this is the same shape as the issue itself: a script that only works where its author's files happen to be. It just showed up in sys.path instead of in an output path.

[wshlavacek]

Closes#489. Follows #490, which fixed
the two runners run_all.py can reach.

The paper side's comment on the issue is followed exactly: _ssa_cell.py and
bngsim_wrapper.py get no parser, that exemption is now endorsed by a test
rather than left to the next implementer, and the freedom granted over
recover_s4_points.py is used.

The hazard was demonstrated, not argued

Two of the seven wrote committed files on a bare invocation. Reproduced
against this repository, then restored from git:

$ cd benchmarks/suites/jacobian && python probe_attach.py --help
--help: attached=None n_func_rxns=None ns=None self_check_failed=None err=missing
WROTE results/attach_probe.json
$ git diff --stat -- results/attach_probe.json
.../jacobian/results/attach_probe.json | 48 +---------------------
1 file changed, 2 insertions(+), 46 deletions(-)

46 lines of measurement replaced by [{"model": "--help", "error": "missing"}] —
by the command typed to find out what the script does. That file is the record of
which models genuinely attach the analytical Jacobian, which is what decides
membership in the suite's speedup table.

script	what --help used to do
jacobian/probe_attach.py	probed a model named --help; overwrote the committedresults/attach_probe.json
ode_engines_s4_sbml/check_sbml_engine_agreement.py	ran bngsim + RoadRunner + COPASI over the Table S4 models, then overwrote its committed report
ode_fullnet/recover_s4_points.py	regenerated two networks through BNG2.pl, rewrote the characterization under parity_checks/, then copied both into another repository'slatex/generated/
ssa_table5/convert_all.py	reconverted all 14 models, rewrote results/converted/ + conversion_log.json
ssa_table5/emit_ssa_table.py	rewrote ssa_timing_ballpark.json in place and re-rendered the .md
ssa_table5/merge_jobout.py	rebuilt ssa_timing_ballpark.json from _jobout/
jacobian/diagnose_divergence.py	read --help as a model id and stopped — the one benign case

The fix

All seven build an ArgumentParser: --help exits 0 with usage, an unrecognized
flag exits 2, a bare invocation is unchanged.

Where a probe needs somewhere harmless to write, it now has a flag — the same
shape --out took on the PSA companion in #490:

• --out on probe_attach.py and check_sbml_engine_agreement.py;
• --paper-dir / --no-copy on recover_s4_points.py. That one also skips
  rather than crashes when the paper checkout is absent: the merge has already
  happened by then, so a FileNotFoundError at the copy left the work done and
  the run marked failed. Its stale /Users/wish/Code/PyBNF/... follow-up line is
  gone too, per your note that it was never a contract.

diagnose_divergence.py is both driver and worker — it re-enters itself as
--worker on an argv it composes in _spawn(). That branch is matched before
the parser and left byte-identical; only the half a person types is parsed.

The exemption, endorsed

ssa_table5/_ssa_cell.py (spawned per (engine, model) by run_ssa_timing.py,
Table 5's measurement path) and sbml_test_suite/testrunner/bngsim_wrapper.py
(invoked by the vendored runner on %d %n %o %l %v) get no parser. Each now
names its caller in a comment saying why there is no parser.

Worth noting: bngsim_wrapper.py was already correct. Its argc check answers
--help with a usage line and exit 2 — the issue's table overstated it as
ignoring argv. It needed the comment, not a change.

The endorsement is a test, not a convention:

deftest_spawned_workers_keep_their_positional_contract(rel):
src= (SUITES_DIR/rel).read_text()
assert"argparse"notinsrc, f"{rel} grew a parser -- it is spawned by ..."assertSPAWNED_WORKERS[rel] insrc, f"{rel} does not name its caller"

So the next argv sweep cannot quietly "fix" these two, and cannot read the
missing parser as an oversight either.

The test now discovers its family

test_benchmark_runner_help.py no longer derives from run_all.py's registry
plus a hand-kept companion list. It globs every suites/ script with a
__main__ guard and subtracts SPAWNED_WORKERS — 59 scripts, 127 cases,
15 s. A script added to any suite inherits the contract without an edit here.
A separate test asserts every registry-named script lands inside that family, so
the orchestrator link stays explicit.

Two tests assert --help leaves each committed artifact byte-identical, which
states the actual harm rather than only "exit 0".

Verification

• 127 cases pass; full local suite green.
• Re-ran every script's --help with libsbml, scipy, roadrunner, amici,
  COPASI, antimony, gillespy2, diffrax, jax, pandas, matplotlib and
  psutil blocked at import. Only two scripts import an optional engine at module
  scope — biomodels/filter.py (libsbml) and showcase/run_ode_trf_fit_from_net.py
  (scipy) — and both deps are guaranteed present wherever this test runs
  (python-libsbml is a core dependency; scipy comes in via the test extra,
  which dev includes and CI syncs).
• diagnose_divergence.py --worker verified to still reach worker() and fail
  inside it on a bogus path, proving the spawn argv unpack is intact.
• The two committed artifacts are byte-identical in this branch.

Risk

None to any measured value. Every script behaves identically when invoked as it
was before, and the two scripts on a measurement path are the two that were not
touched functionally.

[wshlavacek]

CI was red — 19 cases, all in the new test file, all one defect it had just been widened enough to see. Fixed in c031852.

Eight scripts resolved their repo root as Path.home() / "Code" / "bngsim" and then put <root>/parity_checks on sys.path. Their module-scope import _bng_common therefore resolved only on a machine whose clone sits at that path, and raised ModuleNotFoundError in any other checkout — a worktree, a fresh clone, every CI runner:

ode_engines_s3/run_s3_timing.py (+ patch_amici_split.py, which imports it)
ode_engines_s4_sbml/run_s4_timing.py
ode_engines_s4_sbml/check_sbml_engine_agreement.py
ode_fullnet/{run_timing,run_forced,gen_networks,probe_growth,recover_s4_points}.py

The default is now Path(__file__).resolve().parents[3] — the checkout the file lives in — which is the idiom biomodels/emit.py already uses. BNGSIM_ROOT still overrides, so the case the env var was written for (a different venv run against a canonical checkout) is untouched, and where the two already coincided the resolved path is identical.

The test could not have caught it, because the author's $HOME made the wrong default look right. Every probe now runs with HOME/USERPROFILE pointed at an empty session-scoped directory. Stashing one script's fix reproduces the exact CI failure locally, so this class is now caught on a dev machine rather than on a runner.

Worth noting this is the same shape as the issue itself: a script that only works where its author's files happen to be. It just showed up in sys.path instead of in an output path.

[wshlavacek]

Closes#489. Follows #490, which fixed
the two runners run_all.py can reach.

The paper side's comment on the issue is followed exactly: _ssa_cell.py and
bngsim_wrapper.py get no parser, that exemption is now endorsed by a test
rather than left to the next implementer, and the freedom granted over
recover_s4_points.py is used.

The hazard was demonstrated, not argued

Two of the seven wrote committed files on a bare invocation. Reproduced
against this repository, then restored from git:

$ cd benchmarks/suites/jacobian && python probe_attach.py --help
--help: attached=None n_func_rxns=None ns=None self_check_failed=None err=missing
WROTE results/attach_probe.json
$ git diff --stat -- results/attach_probe.json
.../jacobian/results/attach_probe.json | 48 +---------------------
1 file changed, 2 insertions(+), 46 deletions(-)

46 lines of measurement replaced by [{"model": "--help", "error": "missing"}] —
by the command typed to find out what the script does. That file is the record of
which models genuinely attach the analytical Jacobian, which is what decides
membership in the suite's speedup table.

script	what --help used to do
jacobian/probe_attach.py	probed a model named --help; overwrote the committedresults/attach_probe.json
ode_engines_s4_sbml/check_sbml_engine_agreement.py	ran bngsim + RoadRunner + COPASI over the Table S4 models, then overwrote its committed report
ode_fullnet/recover_s4_points.py	regenerated two networks through BNG2.pl, rewrote the characterization under parity_checks/, then copied both into another repository'slatex/generated/
ssa_table5/convert_all.py	reconverted all 14 models, rewrote results/converted/ + conversion_log.json
ssa_table5/emit_ssa_table.py	rewrote ssa_timing_ballpark.json in place and re-rendered the .md
ssa_table5/merge_jobout.py	rebuilt ssa_timing_ballpark.json from _jobout/
jacobian/diagnose_divergence.py	read --help as a model id and stopped — the one benign case

The fix

All seven build an ArgumentParser: --help exits 0 with usage, an unrecognized
flag exits 2, a bare invocation is unchanged.

Where a probe needs somewhere harmless to write, it now has a flag — the same
shape --out took on the PSA companion in #490:

• --out on probe_attach.py and check_sbml_engine_agreement.py;
• --paper-dir / --no-copy on recover_s4_points.py. That one also skips
  rather than crashes when the paper checkout is absent: the merge has already
  happened by then, so a FileNotFoundError at the copy left the work done and
  the run marked failed. Its stale /Users/wish/Code/PyBNF/... follow-up line is
  gone too, per your note that it was never a contract.

diagnose_divergence.py is both driver and worker — it re-enters itself as
--worker on an argv it composes in _spawn(). That branch is matched before
the parser and left byte-identical; only the half a person types is parsed.

The exemption, endorsed

ssa_table5/_ssa_cell.py (spawned per (engine, model) by run_ssa_timing.py,
Table 5's measurement path) and sbml_test_suite/testrunner/bngsim_wrapper.py
(invoked by the vendored runner on %d %n %o %l %v) get no parser. Each now
names its caller in a comment saying why there is no parser.

Worth noting: bngsim_wrapper.py was already correct. Its argc check answers
--help with a usage line and exit 2 — the issue's table overstated it as
ignoring argv. It needed the comment, not a change.

The endorsement is a test, not a convention:

deftest_spawned_workers_keep_their_positional_contract(rel):
src= (SUITES_DIR/rel).read_text()
assert"argparse"notinsrc, f"{rel} grew a parser -- it is spawned by ..."assertSPAWNED_WORKERS[rel] insrc, f"{rel} does not name its caller"

So the next argv sweep cannot quietly "fix" these two, and cannot read the
missing parser as an oversight either.

The test now discovers its family

test_benchmark_runner_help.py no longer derives from run_all.py's registry
plus a hand-kept companion list. It globs every suites/ script with a
__main__ guard and subtracts SPAWNED_WORKERS — 59 scripts, 127 cases,
15 s. A script added to any suite inherits the contract without an edit here.
A separate test asserts every registry-named script lands inside that family, so
the orchestrator link stays explicit.

Two tests assert --help leaves each committed artifact byte-identical, which
states the actual harm rather than only "exit 0".

Verification

• 127 cases pass; full local suite green.
• Re-ran every script's --help with libsbml, scipy, roadrunner, amici,
  COPASI, antimony, gillespy2, diffrax, jax, pandas, matplotlib and
  psutil blocked at import. Only two scripts import an optional engine at module
  scope — biomodels/filter.py (libsbml) and showcase/run_ode_trf_fit_from_net.py
  (scipy) — and both deps are guaranteed present wherever this test runs
  (python-libsbml is a core dependency; scipy comes in via the test extra,
  which dev includes and CI syncs).
• diagnose_divergence.py --worker verified to still reach worker() and fail
  inside it on a bogus path, proving the spawn argv unpack is intact.
• The two committed artifacts are byte-identical in this branch.

Risk

None to any measured value. Every script behaves identically when invoked as it
was before, and the two scripts on a measurement path are the two that were not
touched functionally.

[wshlavacek]

CI was red — 19 cases, all in the new test file, all one defect it had just been widened enough to see. Fixed in c031852.

Eight scripts resolved their repo root as Path.home() / "Code" / "bngsim" and then put <root>/parity_checks on sys.path. Their module-scope import _bng_common therefore resolved only on a machine whose clone sits at that path, and raised ModuleNotFoundError in any other checkout — a worktree, a fresh clone, every CI runner:

ode_engines_s3/run_s3_timing.py (+ patch_amici_split.py, which imports it)
ode_engines_s4_sbml/run_s4_timing.py
ode_engines_s4_sbml/check_sbml_engine_agreement.py
ode_fullnet/{run_timing,run_forced,gen_networks,probe_growth,recover_s4_points}.py

The default is now Path(__file__).resolve().parents[3] — the checkout the file lives in — which is the idiom biomodels/emit.py already uses. BNGSIM_ROOT still overrides, so the case the env var was written for (a different venv run against a canonical checkout) is untouched, and where the two already coincided the resolved path is identical.

The test could not have caught it, because the author's $HOME made the wrong default look right. Every probe now runs with HOME/USERPROFILE pointed at an empty session-scoped directory. Stashing one script's fix reproduces the exact CI failure locally, so this class is now caught on a dev machine rather than on a runner.

Worth noting this is the same shape as the issue itself: a script that only works where its author's files happen to be. It just showed up in sys.path instead of in an output path.

[wshlavacek]

Closes#489. Follows #490, which fixed
the two runners run_all.py can reach.

The paper side's comment on the issue is followed exactly: _ssa_cell.py and
bngsim_wrapper.py get no parser, that exemption is now endorsed by a test
rather than left to the next implementer, and the freedom granted over
recover_s4_points.py is used.

The hazard was demonstrated, not argued

Two of the seven wrote committed files on a bare invocation. Reproduced
against this repository, then restored from git:

$ cd benchmarks/suites/jacobian && python probe_attach.py --help
--help: attached=None n_func_rxns=None ns=None self_check_failed=None err=missing
WROTE results/attach_probe.json
$ git diff --stat -- results/attach_probe.json
.../jacobian/results/attach_probe.json | 48 +---------------------
1 file changed, 2 insertions(+), 46 deletions(-)

46 lines of measurement replaced by [{"model": "--help", "error": "missing"}] —
by the command typed to find out what the script does. That file is the record of
which models genuinely attach the analytical Jacobian, which is what decides
membership in the suite's speedup table.

script	what --help used to do
jacobian/probe_attach.py	probed a model named --help; overwrote the committedresults/attach_probe.json
ode_engines_s4_sbml/check_sbml_engine_agreement.py	ran bngsim + RoadRunner + COPASI over the Table S4 models, then overwrote its committed report
ode_fullnet/recover_s4_points.py	regenerated two networks through BNG2.pl, rewrote the characterization under parity_checks/, then copied both into another repository'slatex/generated/
ssa_table5/convert_all.py	reconverted all 14 models, rewrote results/converted/ + conversion_log.json
ssa_table5/emit_ssa_table.py	rewrote ssa_timing_ballpark.json in place and re-rendered the .md
ssa_table5/merge_jobout.py	rebuilt ssa_timing_ballpark.json from _jobout/
jacobian/diagnose_divergence.py	read --help as a model id and stopped — the one benign case

The fix

All seven build an ArgumentParser: --help exits 0 with usage, an unrecognized
flag exits 2, a bare invocation is unchanged.

Where a probe needs somewhere harmless to write, it now has a flag — the same
shape --out took on the PSA companion in #490:

• --out on probe_attach.py and check_sbml_engine_agreement.py;
• --paper-dir / --no-copy on recover_s4_points.py. That one also skips
  rather than crashes when the paper checkout is absent: the merge has already
  happened by then, so a FileNotFoundError at the copy left the work done and
  the run marked failed. Its stale /Users/wish/Code/PyBNF/... follow-up line is
  gone too, per your note that it was never a contract.

diagnose_divergence.py is both driver and worker — it re-enters itself as
--worker on an argv it composes in _spawn(). That branch is matched before
the parser and left byte-identical; only the half a person types is parsed.

The exemption, endorsed

ssa_table5/_ssa_cell.py (spawned per (engine, model) by run_ssa_timing.py,
Table 5's measurement path) and sbml_test_suite/testrunner/bngsim_wrapper.py
(invoked by the vendored runner on %d %n %o %l %v) get no parser. Each now
names its caller in a comment saying why there is no parser.

Worth noting: bngsim_wrapper.py was already correct. Its argc check answers
--help with a usage line and exit 2 — the issue's table overstated it as
ignoring argv. It needed the comment, not a change.

The endorsement is a test, not a convention:

deftest_spawned_workers_keep_their_positional_contract(rel):
src= (SUITES_DIR/rel).read_text()
assert"argparse"notinsrc, f"{rel} grew a parser -- it is spawned by ..."assertSPAWNED_WORKERS[rel] insrc, f"{rel} does not name its caller"

So the next argv sweep cannot quietly "fix" these two, and cannot read the
missing parser as an oversight either.

The test now discovers its family

test_benchmark_runner_help.py no longer derives from run_all.py's registry
plus a hand-kept companion list. It globs every suites/ script with a
__main__ guard and subtracts SPAWNED_WORKERS — 59 scripts, 127 cases,
15 s. A script added to any suite inherits the contract without an edit here.
A separate test asserts every registry-named script lands inside that family, so
the orchestrator link stays explicit.

Two tests assert --help leaves each committed artifact byte-identical, which
states the actual harm rather than only "exit 0".

Verification

• 127 cases pass; full local suite green.
• Re-ran every script's --help with libsbml, scipy, roadrunner, amici,
  COPASI, antimony, gillespy2, diffrax, jax, pandas, matplotlib and
  psutil blocked at import. Only two scripts import an optional engine at module
  scope — biomodels/filter.py (libsbml) and showcase/run_ode_trf_fit_from_net.py
  (scipy) — and both deps are guaranteed present wherever this test runs
  (python-libsbml is a core dependency; scipy comes in via the test extra,
  which dev includes and CI syncs).
• diagnose_divergence.py --worker verified to still reach worker() and fail
  inside it on a bogus path, proving the spawn argv unpack is intact.
• The two committed artifacts are byte-identical in this branch.

Risk

None to any measured value. Every script behaves identically when invoked as it
was before, and the two scripts on a measurement path are the two that were not
touched functionally.

[wshlavacek]

CI was red — 19 cases, all in the new test file, all one defect it had just been widened enough to see. Fixed in c031852.

Eight scripts resolved their repo root as Path.home() / "Code" / "bngsim" and then put <root>/parity_checks on sys.path. Their module-scope import _bng_common therefore resolved only on a machine whose clone sits at that path, and raised ModuleNotFoundError in any other checkout — a worktree, a fresh clone, every CI runner:

ode_engines_s3/run_s3_timing.py (+ patch_amici_split.py, which imports it)
ode_engines_s4_sbml/run_s4_timing.py
ode_engines_s4_sbml/check_sbml_engine_agreement.py
ode_fullnet/{run_timing,run_forced,gen_networks,probe_growth,recover_s4_points}.py

The default is now Path(__file__).resolve().parents[3] — the checkout the file lives in — which is the idiom biomodels/emit.py already uses. BNGSIM_ROOT still overrides, so the case the env var was written for (a different venv run against a canonical checkout) is untouched, and where the two already coincided the resolved path is identical.

The test could not have caught it, because the author's $HOME made the wrong default look right. Every probe now runs with HOME/USERPROFILE pointed at an empty session-scoped directory. Stashing one script's fix reproduces the exact CI failure locally, so this class is now caught on a dev machine rather than on a runner.

Worth noting this is the same shape as the issue itself: a script that only works where its author's files happen to be. It just showed up in sys.path instead of in an output path.

[wshlavacek]

Closes#489. Follows #490, which fixed
the two runners run_all.py can reach.

The paper side's comment on the issue is followed exactly: _ssa_cell.py and
bngsim_wrapper.py get no parser, that exemption is now endorsed by a test
rather than left to the next implementer, and the freedom granted over
recover_s4_points.py is used.

The hazard was demonstrated, not argued

Two of the seven wrote committed files on a bare invocation. Reproduced
against this repository, then restored from git:

$ cd benchmarks/suites/jacobian && python probe_attach.py --help
--help: attached=None n_func_rxns=None ns=None self_check_failed=None err=missing
WROTE results/attach_probe.json
$ git diff --stat -- results/attach_probe.json
.../jacobian/results/attach_probe.json | 48 +---------------------
1 file changed, 2 insertions(+), 46 deletions(-)

46 lines of measurement replaced by [{"model": "--help", "error": "missing"}] —
by the command typed to find out what the script does. That file is the record of
which models genuinely attach the analytical Jacobian, which is what decides
membership in the suite's speedup table.

script	what --help used to do
jacobian/probe_attach.py	probed a model named --help; overwrote the committedresults/attach_probe.json
ode_engines_s4_sbml/check_sbml_engine_agreement.py	ran bngsim + RoadRunner + COPASI over the Table S4 models, then overwrote its committed report
ode_fullnet/recover_s4_points.py	regenerated two networks through BNG2.pl, rewrote the characterization under parity_checks/, then copied both into another repository'slatex/generated/
ssa_table5/convert_all.py	reconverted all 14 models, rewrote results/converted/ + conversion_log.json
ssa_table5/emit_ssa_table.py	rewrote ssa_timing_ballpark.json in place and re-rendered the .md
ssa_table5/merge_jobout.py	rebuilt ssa_timing_ballpark.json from _jobout/
jacobian/diagnose_divergence.py	read --help as a model id and stopped — the one benign case

The fix

All seven build an ArgumentParser: --help exits 0 with usage, an unrecognized
flag exits 2, a bare invocation is unchanged.

Where a probe needs somewhere harmless to write, it now has a flag — the same
shape --out took on the PSA companion in #490:

• --out on probe_attach.py and check_sbml_engine_agreement.py;
• --paper-dir / --no-copy on recover_s4_points.py. That one also skips
  rather than crashes when the paper checkout is absent: the merge has already
  happened by then, so a FileNotFoundError at the copy left the work done and
  the run marked failed. Its stale /Users/wish/Code/PyBNF/... follow-up line is
  gone too, per your note that it was never a contract.

diagnose_divergence.py is both driver and worker — it re-enters itself as
--worker on an argv it composes in _spawn(). That branch is matched before
the parser and left byte-identical; only the half a person types is parsed.

The exemption, endorsed

ssa_table5/_ssa_cell.py (spawned per (engine, model) by run_ssa_timing.py,
Table 5's measurement path) and sbml_test_suite/testrunner/bngsim_wrapper.py
(invoked by the vendored runner on %d %n %o %l %v) get no parser. Each now
names its caller in a comment saying why there is no parser.

Worth noting: bngsim_wrapper.py was already correct. Its argc check answers
--help with a usage line and exit 2 — the issue's table overstated it as
ignoring argv. It needed the comment, not a change.

The endorsement is a test, not a convention:

deftest_spawned_workers_keep_their_positional_contract(rel):
src= (SUITES_DIR/rel).read_text()
assert"argparse"notinsrc, f"{rel} grew a parser -- it is spawned by ..."assertSPAWNED_WORKERS[rel] insrc, f"{rel} does not name its caller"

So the next argv sweep cannot quietly "fix" these two, and cannot read the
missing parser as an oversight either.

The test now discovers its family

test_benchmark_runner_help.py no longer derives from run_all.py's registry
plus a hand-kept companion list. It globs every suites/ script with a
__main__ guard and subtracts SPAWNED_WORKERS — 59 scripts, 127 cases,
15 s. A script added to any suite inherits the contract without an edit here.
A separate test asserts every registry-named script lands inside that family, so
the orchestrator link stays explicit.

Two tests assert --help leaves each committed artifact byte-identical, which
states the actual harm rather than only "exit 0".

Verification

• 127 cases pass; full local suite green.
• Re-ran every script's --help with libsbml, scipy, roadrunner, amici,
  COPASI, antimony, gillespy2, diffrax, jax, pandas, matplotlib and
  psutil blocked at import. Only two scripts import an optional engine at module
  scope — biomodels/filter.py (libsbml) and showcase/run_ode_trf_fit_from_net.py
  (scipy) — and both deps are guaranteed present wherever this test runs
  (python-libsbml is a core dependency; scipy comes in via the test extra,
  which dev includes and CI syncs).
• diagnose_divergence.py --worker verified to still reach worker() and fail
  inside it on a bogus path, proving the spawn argv unpack is intact.
• The two committed artifacts are byte-identical in this branch.

Risk

None to any measured value. Every script behaves identically when invoked as it
was before, and the two scripts on a measurement path are the two that were not
touched functionally.

[wshlavacek]

CI was red — 19 cases, all in the new test file, all one defect it had just been widened enough to see. Fixed in c031852.

Eight scripts resolved their repo root as Path.home() / "Code" / "bngsim" and then put <root>/parity_checks on sys.path. Their module-scope import _bng_common therefore resolved only on a machine whose clone sits at that path, and raised ModuleNotFoundError in any other checkout — a worktree, a fresh clone, every CI runner:

ode_engines_s3/run_s3_timing.py (+ patch_amici_split.py, which imports it)
ode_engines_s4_sbml/run_s4_timing.py
ode_engines_s4_sbml/check_sbml_engine_agreement.py
ode_fullnet/{run_timing,run_forced,gen_networks,probe_growth,recover_s4_points}.py

The default is now Path(__file__).resolve().parents[3] — the checkout the file lives in — which is the idiom biomodels/emit.py already uses. BNGSIM_ROOT still overrides, so the case the env var was written for (a different venv run against a canonical checkout) is untouched, and where the two already coincided the resolved path is identical.

The test could not have caught it, because the author's $HOME made the wrong default look right. Every probe now runs with HOME/USERPROFILE pointed at an empty session-scoped directory. Stashing one script's fix reproduces the exact CI failure locally, so this class is now caught on a dev machine rather than on a runner.

Worth noting this is the same shape as the issue itself: a script that only works where its author's files happen to be. It just showed up in sys.path instead of in an output path.

[wshlavacek]

Closes#489. Follows #490, which fixed
the two runners run_all.py can reach.

The paper side's comment on the issue is followed exactly: _ssa_cell.py and
bngsim_wrapper.py get no parser, that exemption is now endorsed by a test
rather than left to the next implementer, and the freedom granted over
recover_s4_points.py is used.

The hazard was demonstrated, not argued

Two of the seven wrote committed files on a bare invocation. Reproduced
against this repository, then restored from git:

$ cd benchmarks/suites/jacobian && python probe_attach.py --help
--help: attached=None n_func_rxns=None ns=None self_check_failed=None err=missing
WROTE results/attach_probe.json
$ git diff --stat -- results/attach_probe.json
.../jacobian/results/attach_probe.json | 48 +---------------------
1 file changed, 2 insertions(+), 46 deletions(-)

46 lines of measurement replaced by [{"model": "--help", "error": "missing"}] —
by the command typed to find out what the script does. That file is the record of
which models genuinely attach the analytical Jacobian, which is what decides
membership in the suite's speedup table.

script	what --help used to do
jacobian/probe_attach.py	probed a model named --help; overwrote the committedresults/attach_probe.json
ode_engines_s4_sbml/check_sbml_engine_agreement.py	ran bngsim + RoadRunner + COPASI over the Table S4 models, then overwrote its committed report
ode_fullnet/recover_s4_points.py	regenerated two networks through BNG2.pl, rewrote the characterization under parity_checks/, then copied both into another repository'slatex/generated/
ssa_table5/convert_all.py	reconverted all 14 models, rewrote results/converted/ + conversion_log.json
ssa_table5/emit_ssa_table.py	rewrote ssa_timing_ballpark.json in place and re-rendered the .md
ssa_table5/merge_jobout.py	rebuilt ssa_timing_ballpark.json from _jobout/
jacobian/diagnose_divergence.py	read --help as a model id and stopped — the one benign case

The fix

All seven build an ArgumentParser: --help exits 0 with usage, an unrecognized
flag exits 2, a bare invocation is unchanged.

Where a probe needs somewhere harmless to write, it now has a flag — the same
shape --out took on the PSA companion in #490:

• --out on probe_attach.py and check_sbml_engine_agreement.py;
• --paper-dir / --no-copy on recover_s4_points.py. That one also skips
  rather than crashes when the paper checkout is absent: the merge has already
  happened by then, so a FileNotFoundError at the copy left the work done and
  the run marked failed. Its stale /Users/wish/Code/PyBNF/... follow-up line is
  gone too, per your note that it was never a contract.

diagnose_divergence.py is both driver and worker — it re-enters itself as
--worker on an argv it composes in _spawn(). That branch is matched before
the parser and left byte-identical; only the half a person types is parsed.

The exemption, endorsed

ssa_table5/_ssa_cell.py (spawned per (engine, model) by run_ssa_timing.py,
Table 5's measurement path) and sbml_test_suite/testrunner/bngsim_wrapper.py
(invoked by the vendored runner on %d %n %o %l %v) get no parser. Each now
names its caller in a comment saying why there is no parser.

Worth noting: bngsim_wrapper.py was already correct. Its argc check answers
--help with a usage line and exit 2 — the issue's table overstated it as
ignoring argv. It needed the comment, not a change.

The endorsement is a test, not a convention:

deftest_spawned_workers_keep_their_positional_contract(rel):
src= (SUITES_DIR/rel).read_text()
assert"argparse"notinsrc, f"{rel} grew a parser -- it is spawned by ..."assertSPAWNED_WORKERS[rel] insrc, f"{rel} does not name its caller"

So the next argv sweep cannot quietly "fix" these two, and cannot read the
missing parser as an oversight either.

The test now discovers its family

test_benchmark_runner_help.py no longer derives from run_all.py's registry
plus a hand-kept companion list. It globs every suites/ script with a
__main__ guard and subtracts SPAWNED_WORKERS — 59 scripts, 127 cases,
15 s. A script added to any suite inherits the contract without an edit here.
A separate test asserts every registry-named script lands inside that family, so
the orchestrator link stays explicit.

Two tests assert --help leaves each committed artifact byte-identical, which
states the actual harm rather than only "exit 0".

Verification

• 127 cases pass; full local suite green.
• Re-ran every script's --help with libsbml, scipy, roadrunner, amici,
  COPASI, antimony, gillespy2, diffrax, jax, pandas, matplotlib and
  psutil blocked at import. Only two scripts import an optional engine at module
  scope — biomodels/filter.py (libsbml) and showcase/run_ode_trf_fit_from_net.py
  (scipy) — and both deps are guaranteed present wherever this test runs
  (python-libsbml is a core dependency; scipy comes in via the test extra,
  which dev includes and CI syncs).
• diagnose_divergence.py --worker verified to still reach worker() and fail
  inside it on a bogus path, proving the spawn argv unpack is intact.
• The two committed artifacts are byte-identical in this branch.

Risk

None to any measured value. Every script behaves identically when invoked as it
was before, and the two scripts on a measurement path are the two that were not
touched functionally.

[wshlavacek]

CI was red — 19 cases, all in the new test file, all one defect it had just been widened enough to see. Fixed in c031852.

Eight scripts resolved their repo root as Path.home() / "Code" / "bngsim" and then put <root>/parity_checks on sys.path. Their module-scope import _bng_common therefore resolved only on a machine whose clone sits at that path, and raised ModuleNotFoundError in any other checkout — a worktree, a fresh clone, every CI runner:

ode_engines_s3/run_s3_timing.py (+ patch_amici_split.py, which imports it)
ode_engines_s4_sbml/run_s4_timing.py
ode_engines_s4_sbml/check_sbml_engine_agreement.py
ode_fullnet/{run_timing,run_forced,gen_networks,probe_growth,recover_s4_points}.py

The default is now Path(__file__).resolve().parents[3] — the checkout the file lives in — which is the idiom biomodels/emit.py already uses. BNGSIM_ROOT still overrides, so the case the env var was written for (a different venv run against a canonical checkout) is untouched, and where the two already coincided the resolved path is identical.

The test could not have caught it, because the author's $HOME made the wrong default look right. Every probe now runs with HOME/USERPROFILE pointed at an empty session-scoped directory. Stashing one script's fix reproduces the exact CI failure locally, so this class is now caught on a dev machine rather than on a runner.

Worth noting this is the same shape as the issue itself: a script that only works where its author's files happen to be. It just showed up in sys.path instead of in an output path.

[wshlavacek]

Closes#489. Follows #490, which fixed
the two runners run_all.py can reach.

The paper side's comment on the issue is followed exactly: _ssa_cell.py and
bngsim_wrapper.py get no parser, that exemption is now endorsed by a test
rather than left to the next implementer, and the freedom granted over
recover_s4_points.py is used.

The hazard was demonstrated, not argued

Two of the seven wrote committed files on a bare invocation. Reproduced
against this repository, then restored from git:

$ cd benchmarks/suites/jacobian && python probe_attach.py --help
--help: attached=None n_func_rxns=None ns=None self_check_failed=None err=missing
WROTE results/attach_probe.json
$ git diff --stat -- results/attach_probe.json
.../jacobian/results/attach_probe.json | 48 +---------------------
1 file changed, 2 insertions(+), 46 deletions(-)

46 lines of measurement replaced by [{"model": "--help", "error": "missing"}] —
by the command typed to find out what the script does. That file is the record of
which models genuinely attach the analytical Jacobian, which is what decides
membership in the suite's speedup table.

script	what --help used to do
jacobian/probe_attach.py	probed a model named --help; overwrote the committedresults/attach_probe.json
ode_engines_s4_sbml/check_sbml_engine_agreement.py	ran bngsim + RoadRunner + COPASI over the Table S4 models, then overwrote its committed report
ode_fullnet/recover_s4_points.py	regenerated two networks through BNG2.pl, rewrote the characterization under parity_checks/, then copied both into another repository'slatex/generated/
ssa_table5/convert_all.py	reconverted all 14 models, rewrote results/converted/ + conversion_log.json
ssa_table5/emit_ssa_table.py	rewrote ssa_timing_ballpark.json in place and re-rendered the .md
ssa_table5/merge_jobout.py	rebuilt ssa_timing_ballpark.json from _jobout/
jacobian/diagnose_divergence.py	read --help as a model id and stopped — the one benign case

The fix

All seven build an ArgumentParser: --help exits 0 with usage, an unrecognized
flag exits 2, a bare invocation is unchanged.

Where a probe needs somewhere harmless to write, it now has a flag — the same
shape --out took on the PSA companion in #490:

• --out on probe_attach.py and check_sbml_engine_agreement.py;
• --paper-dir / --no-copy on recover_s4_points.py. That one also skips
  rather than crashes when the paper checkout is absent: the merge has already
  happened by then, so a FileNotFoundError at the copy left the work done and
  the run marked failed. Its stale /Users/wish/Code/PyBNF/... follow-up line is
  gone too, per your note that it was never a contract.

diagnose_divergence.py is both driver and worker — it re-enters itself as
--worker on an argv it composes in _spawn(). That branch is matched before
the parser and left byte-identical; only the half a person types is parsed.

The exemption, endorsed

ssa_table5/_ssa_cell.py (spawned per (engine, model) by run_ssa_timing.py,
Table 5's measurement path) and sbml_test_suite/testrunner/bngsim_wrapper.py
(invoked by the vendored runner on %d %n %o %l %v) get no parser. Each now
names its caller in a comment saying why there is no parser.

Worth noting: bngsim_wrapper.py was already correct. Its argc check answers
--help with a usage line and exit 2 — the issue's table overstated it as
ignoring argv. It needed the comment, not a change.

The endorsement is a test, not a convention:

deftest_spawned_workers_keep_their_positional_contract(rel):
src= (SUITES_DIR/rel).read_text()
assert"argparse"notinsrc, f"{rel} grew a parser -- it is spawned by ..."assertSPAWNED_WORKERS[rel] insrc, f"{rel} does not name its caller"

So the next argv sweep cannot quietly "fix" these two, and cannot read the
missing parser as an oversight either.

The test now discovers its family

test_benchmark_runner_help.py no longer derives from run_all.py's registry
plus a hand-kept companion list. It globs every suites/ script with a
__main__ guard and subtracts SPAWNED_WORKERS — 59 scripts, 127 cases,
15 s. A script added to any suite inherits the contract without an edit here.
A separate test asserts every registry-named script lands inside that family, so
the orchestrator link stays explicit.

Two tests assert --help leaves each committed artifact byte-identical, which
states the actual harm rather than only "exit 0".

Verification

• 127 cases pass; full local suite green.
• Re-ran every script's --help with libsbml, scipy, roadrunner, amici,
  COPASI, antimony, gillespy2, diffrax, jax, pandas, matplotlib and
  psutil blocked at import. Only two scripts import an optional engine at module
  scope — biomodels/filter.py (libsbml) and showcase/run_ode_trf_fit_from_net.py
  (scipy) — and both deps are guaranteed present wherever this test runs
  (python-libsbml is a core dependency; scipy comes in via the test extra,
  which dev includes and CI syncs).
• diagnose_divergence.py --worker verified to still reach worker() and fail
  inside it on a bogus path, proving the spawn argv unpack is intact.
• The two committed artifacts are byte-identical in this branch.

Risk

None to any measured value. Every script behaves identically when invoked as it
was before, and the two scripts on a measurement path are the two that were not
touched functionally.

[wshlavacek]

CI was red — 19 cases, all in the new test file, all one defect it had just been widened enough to see. Fixed in c031852.

Eight scripts resolved their repo root as Path.home() / "Code" / "bngsim" and then put <root>/parity_checks on sys.path. Their module-scope import _bng_common therefore resolved only on a machine whose clone sits at that path, and raised ModuleNotFoundError in any other checkout — a worktree, a fresh clone, every CI runner:

ode_engines_s3/run_s3_timing.py (+ patch_amici_split.py, which imports it)
ode_engines_s4_sbml/run_s4_timing.py
ode_engines_s4_sbml/check_sbml_engine_agreement.py
ode_fullnet/{run_timing,run_forced,gen_networks,probe_growth,recover_s4_points}.py

The default is now Path(__file__).resolve().parents[3] — the checkout the file lives in — which is the idiom biomodels/emit.py already uses. BNGSIM_ROOT still overrides, so the case the env var was written for (a different venv run against a canonical checkout) is untouched, and where the two already coincided the resolved path is identical.

The test could not have caught it, because the author's $HOME made the wrong default look right. Every probe now runs with HOME/USERPROFILE pointed at an empty session-scoped directory. Stashing one script's fix reproduces the exact CI failure locally, so this class is now caught on a dev machine rather than on a runner.

Worth noting this is the same shape as the issue itself: a script that only works where its author's files happen to be. It just showed up in sys.path instead of in an output path.

[wshlavacek]

Closes#489. Follows #490, which fixed
the two runners run_all.py can reach.

The paper side's comment on the issue is followed exactly: _ssa_cell.py and
bngsim_wrapper.py get no parser, that exemption is now endorsed by a test
rather than left to the next implementer, and the freedom granted over
recover_s4_points.py is used.

The hazard was demonstrated, not argued

Two of the seven wrote committed files on a bare invocation. Reproduced
against this repository, then restored from git:

$ cd benchmarks/suites/jacobian && python probe_attach.py --help
--help: attached=None n_func_rxns=None ns=None self_check_failed=None err=missing
WROTE results/attach_probe.json
$ git diff --stat -- results/attach_probe.json
.../jacobian/results/attach_probe.json | 48 +---------------------
1 file changed, 2 insertions(+), 46 deletions(-)

46 lines of measurement replaced by [{"model": "--help", "error": "missing"}] —
by the command typed to find out what the script does. That file is the record of
which models genuinely attach the analytical Jacobian, which is what decides
membership in the suite's speedup table.

script	what --help used to do
jacobian/probe_attach.py	probed a model named --help; overwrote the committedresults/attach_probe.json
ode_engines_s4_sbml/check_sbml_engine_agreement.py	ran bngsim + RoadRunner + COPASI over the Table S4 models, then overwrote its committed report
ode_fullnet/recover_s4_points.py	regenerated two networks through BNG2.pl, rewrote the characterization under parity_checks/, then copied both into another repository'slatex/generated/
ssa_table5/convert_all.py	reconverted all 14 models, rewrote results/converted/ + conversion_log.json
ssa_table5/emit_ssa_table.py	rewrote ssa_timing_ballpark.json in place and re-rendered the .md
ssa_table5/merge_jobout.py	rebuilt ssa_timing_ballpark.json from _jobout/
jacobian/diagnose_divergence.py	read --help as a model id and stopped — the one benign case

The fix

All seven build an ArgumentParser: --help exits 0 with usage, an unrecognized
flag exits 2, a bare invocation is unchanged.

Where a probe needs somewhere harmless to write, it now has a flag — the same
shape --out took on the PSA companion in #490:

• --out on probe_attach.py and check_sbml_engine_agreement.py;
• --paper-dir / --no-copy on recover_s4_points.py. That one also skips
  rather than crashes when the paper checkout is absent: the merge has already
  happened by then, so a FileNotFoundError at the copy left the work done and
  the run marked failed. Its stale /Users/wish/Code/PyBNF/... follow-up line is
  gone too, per your note that it was never a contract.

diagnose_divergence.py is both driver and worker — it re-enters itself as
--worker on an argv it composes in _spawn(). That branch is matched before
the parser and left byte-identical; only the half a person types is parsed.

The exemption, endorsed

ssa_table5/_ssa_cell.py (spawned per (engine, model) by run_ssa_timing.py,
Table 5's measurement path) and sbml_test_suite/testrunner/bngsim_wrapper.py
(invoked by the vendored runner on %d %n %o %l %v) get no parser. Each now
names its caller in a comment saying why there is no parser.

Worth noting: bngsim_wrapper.py was already correct. Its argc check answers
--help with a usage line and exit 2 — the issue's table overstated it as
ignoring argv. It needed the comment, not a change.

The endorsement is a test, not a convention:

deftest_spawned_workers_keep_their_positional_contract(rel):
src= (SUITES_DIR/rel).read_text()
assert"argparse"notinsrc, f"{rel} grew a parser -- it is spawned by ..."assertSPAWNED_WORKERS[rel] insrc, f"{rel} does not name its caller"

So the next argv sweep cannot quietly "fix" these two, and cannot read the
missing parser as an oversight either.

The test now discovers its family

test_benchmark_runner_help.py no longer derives from run_all.py's registry
plus a hand-kept companion list. It globs every suites/ script with a
__main__ guard and subtracts SPAWNED_WORKERS — 59 scripts, 127 cases,
15 s. A script added to any suite inherits the contract without an edit here.
A separate test asserts every registry-named script lands inside that family, so
the orchestrator link stays explicit.

Two tests assert --help leaves each committed artifact byte-identical, which
states the actual harm rather than only "exit 0".

Verification

• 127 cases pass; full local suite green.
• Re-ran every script's --help with libsbml, scipy, roadrunner, amici,
  COPASI, antimony, gillespy2, diffrax, jax, pandas, matplotlib and
  psutil blocked at import. Only two scripts import an optional engine at module
  scope — biomodels/filter.py (libsbml) and showcase/run_ode_trf_fit_from_net.py
  (scipy) — and both deps are guaranteed present wherever this test runs
  (python-libsbml is a core dependency; scipy comes in via the test extra,
  which dev includes and CI syncs).
• diagnose_divergence.py --worker verified to still reach worker() and fail
  inside it on a bogus path, proving the spawn argv unpack is intact.
• The two committed artifacts are byte-identical in this branch.

Risk

None to any measured value. Every script behaves identically when invoked as it
was before, and the two scripts on a measurement path are the two that were not
touched functionally.

[wshlavacek]

CI was red — 19 cases, all in the new test file, all one defect it had just been widened enough to see. Fixed in c031852.

Eight scripts resolved their repo root as Path.home() / "Code" / "bngsim" and then put <root>/parity_checks on sys.path. Their module-scope import _bng_common therefore resolved only on a machine whose clone sits at that path, and raised ModuleNotFoundError in any other checkout — a worktree, a fresh clone, every CI runner:

ode_engines_s3/run_s3_timing.py (+ patch_amici_split.py, which imports it)
ode_engines_s4_sbml/run_s4_timing.py
ode_engines_s4_sbml/check_sbml_engine_agreement.py
ode_fullnet/{run_timing,run_forced,gen_networks,probe_growth,recover_s4_points}.py

The default is now Path(__file__).resolve().parents[3] — the checkout the file lives in — which is the idiom biomodels/emit.py already uses. BNGSIM_ROOT still overrides, so the case the env var was written for (a different venv run against a canonical checkout) is untouched, and where the two already coincided the resolved path is identical.

The test could not have caught it, because the author's $HOME made the wrong default look right. Every probe now runs with HOME/USERPROFILE pointed at an empty session-scoped directory. Stashing one script's fix reproduces the exact CI failure locally, so this class is now caught on a dev machine rather than on a runner.

Worth noting this is the same shape as the issue itself: a script that only works where its author's files happen to be. It just showed up in sys.path instead of in an output path.