feat: add binary classification and proper-scoring alert metrics

[knutdrand]

Phase 2 of the alerting track. Stacked on #581 → #580 → #539. Last one in the stack; it retargets down as each merges.

The metrics

Scoring the alert decision:

Metric	Objective?	Why
outbreak_precision	no	maximised by alerting once, and only when certain
false_alarm_rate	no	maximised by never alerting
outbreak_f1	maximize	harmonic mean of precision and sensitivity — prices both errors
outbreak_mcc	maximize	uses all four confusion cells, so true negatives count

Scoring the probability behind it:

Metric	Objective?
outbreak_brier	minimize	mean squared error of the alert probability
outbreak_log_score	minimize	punishes confident misses hard; clipped at 1e-6 so one costs a large finite amount
outbreak_brier_skill	maximize	the number to lead with

Brier skill is the headline because outbreaks are rare: a raw score flatters any model that rarely alerts, while skill prices it against the climatological base rate instead of against nothing.

optimization_direction is set only where both error types are priced. Precision and false-alarm rate join sensitivity and specificity in leaving it unset, for the same reason #580 unset those.

A correction to the plan: GlobalOnlyMetric

The design note claimed this set would fit the existing compute_detailed → mean/sum/RMS contract, and that only AUROC would not. That was wrong. F1, Matthews correlation and Brier skill are all ratios of aggregates — there is no per-cell value whose mean recovers them, exactly the property that rules AUROC out.

Deferring all three would have left Phase 2 as precision, false-alarm rate, Brier and log score: the metrics nobody should lead with, minus every one that prices both errors. So this adds GlobalOnlyMetric to metrics/base.py — ~35 lines that report one value and raise when asked for a breakdown, rather than returning a number that looks per-cell and is not. AUROC now needs no new machinery whenever we want it.

One product gap this surfaced

test_generate_metric_visualization parametrizes over every metric the plot picker offers, and three failed: every metric plot breaks a score down by horizon, location or period, which a global-only metric has no values for. That was not a test artifact — the picker was offering metrics that render an error instead of a chart.

GET /metric-plots's metric list now filters to metrics with a per-cell value, with a regression test. Global-only scores are read from the backtest's aggregate metrics instead.

Testing

• make lint (ruff, mypy, pyright) passes
• uv run pytest -q --ignore=.claude: 1526 passed, 122 skipped, 4 xfailed, 1 xpassed
• 16 new tests. The balanced_alert_scenario fixture yields exactly one TP, FP, TN and FN, so every confusion-matrix metric has a hand-checkable value — precision 0.5, F1 0.5, MCC 0.0 (chance agreement), Brier 0.5, Brier skill −1.0 (worse than climatology). Plus the log-score clip, both global-only refusals, the undefined-skill case, the objective lock-in, and the picker filter.

Still deferred

AUROC and AUPRC (machinery now exists), lead time, and skill against the persistence baseline from #581 — that last one needs last_seen_period plumbed through the flat forecast representation, the same prerequisite as the leakage cutoff, so both belong to that change.