Fix two unsatisfiable dependency floors, make the floors job real, and audit the plan's enforcement claims

[sgjholt]

Completes the floors CI change already applied to main. Main's floors job is red until this merges — it now calls tools/check_floors.py, which lands here.

The job has never tested the floors

- run: uv pip install --resolution lowest-direct -e ".[dev]"
- run: uv run pytest -m "not dataset" -q

uv run re-resolves the project before running. With no committed lock file it installed the newest of everything and discarded the line above it, so the job was a duplicate of the ubuntu/3.11 matrix entry — green for that reason, for as long as it has existed.

Its own comment read "Exercise the declared minimums. CI otherwise installs the newest of everything, so the floors are only ever tested by a user", describing precisely what continued to happen.

Reproduced rather than inferred: uv run prints Uninstalled 2 packages / Installed 2 packages and comes back with lmfit 1.3.4 and numpy 2.4.6 where the previous step had put 1.2.0 and 2.0.0.

Two declared floors were wrong

Both broke real behaviour, and each was bisected rather than guessed.

declared	measured	what fails below it
lmfit>=1.2	>=1.3	1.2.x calls np.asfarray, removed in NumPy 2.0. lmfit>=1.2 and numpy>=2.0 were never both satisfiable; every fit raised AttributeError and 15 tests fail. 1.2.0, 1.2.1, 1.2.2 all fail; 1.3.0 works.
scipy>=1.13	>=1.15	1.15 reimplemented scipy.optimize.nnls. Below it the vendored qiinv inversion does not converge for every input scale, and the quadratic estimator's peak recovery moves between 0.53 and 1.02 for the same signal at different amplitudes. 1.13 and 1.14 fail; 1.15 passes.

The scipy one is the better illustration. test_quadratic.py already asserts scale invariance, with the message "a dimensionless ratio that moves with amplitude is a numerical-range problem, not a property of the estimator". It was right, and it had never run against the floor.

Both raised to what the suite was measured to pass on: 523 pass at the corrected floors, through the fixed job end to end.

The guard matters more than the floors

tools/check_floors.py asserts the installed versions are the declared minimums, and the job runs it before pytest. This is the point of the change rather than a nicety: a floors job quietly testing the newest versions looks exactly like one that works, so the mismatch has to be an error rather than something a reader might notice.

Verified both directions — it passes against a floor environment and fails against the dev environment — and without --no-sync uv re-syncs and the guard turns the job red.

Its first version was itself wrong: it called 2.0 and 2.0.0 different versions and failed against everything. Both sides are zero-padded to equal length now, not truncated to the shorter, which would have accepted 2.0.5 for a floor of 2.0.

uv.lock is gitignored. Not committing one is deliberate — a lock file is exactly what would make --resolution lowest-direct meaningless again.

A stub that over-promised, and the check that missed it

stubs/lmfit/model.pyi declared Model.fit(..., coerce_farray=...), which exists on lmfit 1.3 and not on 1.2. Removed — a stub has to be true across the supported range, and nothing here passes it.

The more useful half is why tests/test_stubs.py did not catch it. It skipped any function whose real signature ends in *args/**kwargs, reasoning that a stub may stop early there. That reasoning is right, and is the whole approach — declare what we use, no more. But may stop early is not may invent, and **kwargs upstream makes an invented parameter worse: the call is swallowed at runtime, so the mistake surfaces as a silently ignored argument rather than a TypeError.

The check now separates the two claims — every declared name must exist, and the ones that do must be in order — through one helper shared by the ObsPy and lmfit tests instead of two near-copies. Verified both ways: restoring coerce_farray passes against installed lmfit 1.3.4 and fails against 1.2.0 naming the parameter. The old check passed against both.

Audit of the plan's enforcement claims (new §6.6)

This is the third time a claim in REFACTOR_PLAN.md has described a mechanism that did not exist — after "CI can assert it never grows" for the mypy backlog, and the ignore_missing_imports recommendation that would have left every ObsPy and lmfit object as Any. One shape each time: an intention stated in the plan, restated as fact in a comment, with nothing behind it.

So the rest were checked rather than read. Four more were false, and are corrected in place so the document does not have to be read alongside its own errata:

Claim	Reality
"pre-commit runs ruff, mypy, nbstripout, check-added-large-files"	mypy is not a hook and never was. Defensible as a decision — it needs the project environment, and the typecheck job covers it — but not a hook that exists.
"Conventional Commits, commitlint-enforced"	No such hook. Followed by hand.
"sphinx-build -W … an undocumented public symbol fails the build"	Wrong even once the docs build exists. -W promotes warnings Sphinx already emits, and autodoc emits none for a symbol it was never asked to document. Needs sphinx.ext.coverage or nitpicky.
test.yml = one matrix doing ruff, mypy, pytest	Five jobs: lint, typecheck, test, floors, plus build.yml.

One case that is neither true nor false, and is filed separately for that reason. The plan says "regression tests pin an explicit config file, never the defaults, so changing a default cannot silently move a golden test." Nothing pins a config — the golden tests read the shipped defaults. But the property was tested rather than the mechanism: bumping smoothing.n_bins from 151 to 158 makes 9 of 25 golden tests fail, loudly, because the golden values are frozen in a committed JSON file. The protection exists by a different route than claimed. The residual gap is real but narrower: a regenerated reference silently adopts whatever defaults were current, since the file records no config.

Three claims checked and holding: the Parseval contract parametrised over ESTIMATORS, PLOT_COLUMNS having exactly one definition, and the registry-wide noise-rotation property test.

One passage is simply stale: §5.2 still describes BW_METHOD/ROT_METHOD as globals in spectral, "a shell over core", with sp.ROT_METHOD = 1 as a live escape hatch. Both spectral.py and _config_legacy.py were deleted in phase 2.

The audit also surfaced something in the other direction: the pre-commit config has a local no-session-links commit-msg hook that appeared nowhere in the plan, and it is the only mechanism keeping private session URLs out of a public repository's history. Now written down.

Adding §6.6 moved branch layout to §6.7. All 21 §6.x cross-references were re-resolved against the headings programmatically rather than by eye; four pointed at the wrong section afterwards and are fixed.

Checks

• All 10 CI jobs green on the head commit, floors included
• Floors job replayed locally end to end: all 16 declared floors are the installed versions, then 523 passed, 5 skipped
• Dev environment: 543 passed, 1 skipped; ruff check, ruff format --check, mypy clean