Make timing_source part of the PsychophysicalKernel key, drop mixed, rename nidaq to nidq

[rojasgabriel]

Goal

Promote timing_source from a dependent attribute to part of the PsychophysicalKernel primary key, remove the mixed value, and rename nidaq to nidq.

A kernel row is one fit from one clock. Comparing NIDQ timing against Bpod timing is then a query over two rows, not a stored third row.

Today timing_source sits below the --- divider, so only one row can exist per analysis set x subject x condition x fit config, and it carries whichever clock happened to populate it. The same fit config under two clocks collides instead of coexisting, and mixed exists as a workaround for not being able to store the pair.

Plan

1. labdata_plugin/analysisschema.py - move timing_source above the --- divider in PsychophysicalKernel, so the key becomes BehaviorAnalysisSet x Subject x trialset_description x PsychophysicalKernelFitConfig x timing_source. Set the enum to exactly enum('nidq', 'bpod').
2. labdata_plugin/analysisschema.py - add timing_source to key_source so populate produces one row per available timing source rather than whichever one the loader returned.
3. src/behavior_analyses/kernel_timing.py - fetch_pooled_kernel_inputs takes the requested timing source as an argument and raises when the selected trialsets cannot supply it. It must not infer a source, and must not silently pool across sources. Update the value it returns from nidaq to nidq.
4. Grep the repo for nidaq and update the remaining references.
5. Add a test that fits one session under both timing sources and asserts two distinct rows sharing a fit config.

Do not touch: the fit math in src/behavior_analyses/kernels.py, the PsychophysicalKernelFitConfig contents, or the psychometric tables. Do not add any table, attribute, or row that stores a NIDQ-versus-Bpod difference; that is a query, not a measurement.

Validation

uvx ruff check .
uvx ruff format --check .

• Run the test suite under Python 3.11. Python 3.10 hits the known SciPy _spropack import problem locally.
• Populate one GRB006 session under both timing sources and confirm two rows differing only in timing_source.

Expect: two rows, clean lint, tests pass.

Done when

• timing_source is part of the primary key
• The enum is exactly nidq and bpod
• Requesting a timing source the trialsets cannot supply raises instead of falling back
• The same fit config can be stored under both timing sources for one session
• A test covers the two-row case

Notes

• This is a key change, so the table must be dropped and re-declared. Check the live row count and any existing mixed rows first. The nidaq to nidq rename comes free with the re-declaration.
• Populating requires the lab DataJoint connection, so the live validation step cannot run in a cloud agent.
• Context: ephys PR #35 removed the duplicate session-level kernel on 2026-08-03, making this table the single canonical kernel surface.