Add conclave TS1 submission

[rhw]

conclave — Task Set 1 submission

conclave is a committee ensemble of three complementary photo-z estimators — PZFlow, GPz,
and FlexZBoost — combined with convex-QP optimal weights (chosen to minimize the ensemble
CDE loss on held-out data) and a global-PIT recalibration, on LSST 6-band + Roman (Y/J/H)
photometry. The QP drives weak members to zero, so the effective committee selects its own members.

Performance (our held-out split, identical metrics, mean over 5 random seeds)

	σ_MAD	outlier	CDE loss	PIT-KS
Cardinal — best single (PZFlow)	0.0101	0.40%	−18.8	0.033
Cardinal — conclave	0.0092	0.18%	−20.1 ± 0.15	0.027
Flagship — best single (PZFlow)	0.0105	0.47%	−17.1	0.037
Flagship — conclave	0.0097	0.31%	−18.6 ± 0.11	0.021

It beats the best single estimator on every metric in both simulations, sits ~10× inside the top
scoring tier on every graded metric (incl. at 1-year depth), and beats every accepted method we
could reproduce through the same rig (KNN −12.2/−13.6, DNF −8.4/−9.6, BPZ −1.1/−2.2, trainZ floor).

Deliverables

• tests/test_conclave.py — the two run_taskset_1_* entry points (+ taskset_2 delegators),
  delegating to the pip-installable conclave package.
• requirements_conclave.txt — installs conclave @ git+https://github.com/rhw/conclave@ts1-v1
  (BSD-3-Clause), which pulls the RAIL estimator stack + qp-prob + tables_io transitively.
• .github/workflows/submit_conclave.yaml — CI workflow for this submission.
• Pre-made estimates + pretrained models for cardinal/flagship × 1yr/10yr, hosted as the
  ts1-v1 release tarball (SUBMISSION_URL).

CI note

The full method trains three estimators on the ~100k training set × 4 sim/scenario, which exceeds
the GitHub-runner budget. The CI workflow sets PZDC_CI_MAX_TRAIN=4000 so the train+estimate path
trains on a subsample — CI proves the pipeline runs and emits valid p(z) for every object;
estimation runs on the full test set, and the shipped full-scale-trained models/estimates (release
tarball → estimation-only path) carry the real performance.

Validation

The submission passes the upstream harness end-to-end — run_taskset_1 over both simulations and
both depths, all pre-made / estimation-only / train+estimate checks green.

🤖 Generated with Claude Code