fix: flux defaults to KL per eq 7; unify four drifted defaults

[nceglia]

Two small fixes and one new guard.

1. phenotypic_flux defaults to KL

METRICS eq 7 defines phenotypic flux as the KL divergence. The code defaulted to
l1. Now distance_metric="kl"; l1 and jsd remain available as extensions.

⚠️ This changes reported flux numbers for any caller relying on the default. The
kernels already existed in tcri/_distance.py — only the default moved. Retires
OPEN_QUESTIONS['flux_distance_default'].

2. Four defaults declared twice, with different values

A knob declared in both layers means the outer declaration silently wins and the inner
one is dead — reachable only by constructing TCRIModule or UnifiedTrainingPlan
directly, which is exactly what a fixture or downstream user does.

knob	outer	inner	effective	now
global_scale (α)	TCRIModel 5.0	TCRIModule 10.0	5.0	5.0
local_scale (β)	TCRIModel 3.0	TCRIModule 5.0	3.0	3.0
n_steps_kl_warmup	train() 2000	UnifiedTrainingPlan 1000	2000	2000
reconstruction_loss_scale	—	—	1e-3	unified earlier in 19db68e

Each is unified to the value that was already effective, so no fitted result
changes. This is not a choice about the right hyperparameter — it removes a value that
silently does nothing.

3. tests/test_shared_defaults.py — guards the class, not the instances

Compares every shared default across the two construction seams
(TCRIModel.__init__ → TCRIModule.__init__, TCRIModel.train →
UnifiedTrainingPlan), plus a regression lock on the four known ones.

global_scale was found by this test on its first run, having been missed by every
manual pass including the one earlier today that turned up the other three.

The existing knob test could not catch this: it verifies a value arrives at its target,
which it does. The defect is that the two declared defaults differ, so anyone reading
one signature is misled about what the package does. Signature-only — no model is
constructed, runs in ~3 s.

Tests

177 passed, 3 skipped.

🤖 Generated with Claude Code

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.