perf(db): tune autovacuum per-table + one-shot ANALYZE helper

Two issues spotted in pg_stat_user_tables:

1. Most `hrrr_profiles` partitions have `last_autoanalyze IS NULL`.
   The partitions are append-only from the hourly grid worker, so
   autovacuum never crosses the analyze threshold — but without
   stats the planner falls back to estimates and picks bad joins
   on per-QSO lookups.

2. `contacts`, `grid_tasks`, `hrrr_fetch_tasks`, and
   `contact_common_volume_radar` see high UPDATE volumes
   (enrichment status flips, queue claim/complete cycles, mechanism
   classifier writes) but only accrue ~5% dead before the default
   20% / 10% autovacuum triggers fire. BackfillEnqueueWorker's
   status-priority ORDER BY + the Rust workers' FOR UPDATE SKIP
   LOCKED claim scan both want fresher stats than that.

Changes:
- Migration `20260424204656_tune_autovacuum_for_high_churn` sets
  per-table `autovacuum_*_scale_factor` + `_threshold = 50` on the
  four tables above. Aggressive enough to keep stats current;
  tame enough that vacuum doesn't thrash the Turing Pi 2 Postgres
  node. Reversible.
- `Weather.analyze_all/1` walks `pg_stat_user_tables`, runs
  `ANALYZE` on every table not auto-analyzed in the last 6 h,
  reports counts. Meant to be invoked once via
  `bin/microwaveprop rpc 'Microwaveprop.Weather.analyze_all()'`
  after deploy to seed statistics on the cold partitions;
  idempotent on re-run.

lib/microwaveprop/weather.ex

@@ -410,6 +410,59 @@ defmodule Microwaveprop.Weather do
     {:ok, n}
   end
 
+  @doc """
+  Run `ANALYZE` on every public-schema table that has not been auto-
+  analyzed recently. Callable from a release shell:
+
+      bin/microwaveprop rpc 'Microwaveprop.Weather.analyze_all()'
+
+  Surfaces the HRRR partition problem: most partitions have zero
+  analyze stats, so the planner picks nested-loop joins that scan
+  81 M rows. Running ANALYZE once gives it live_tup and column
+  histograms to work with.
+
+  `skip_recent` skips anything auto-analyzed in the last N seconds
+  (default 6 h) so re-running is cheap.
+  """
+  @spec analyze_all(keyword()) :: %{analyzed: non_neg_integer(), skipped: non_neg_integer()}
+  def analyze_all(opts \\ []) do
+    skip_recent = Keyword.get(opts, :skip_recent_seconds, 6 * 3600)
+
+    q = """
+    SELECT relname
+    FROM pg_stat_user_tables
+    WHERE schemaname = 'public'
+      AND (last_autoanalyze IS NULL
+           OR last_autoanalyze < now() - ($1 || ' seconds')::interval)
+    """
+
+    %{rows: rows} = Repo.query!(q, [Integer.to_string(skip_recent)])
+    tables = Enum.map(rows, fn [n] -> n end)
+
+    Enum.each(tables, fn t ->
+      Repo.query!("ANALYZE #{quote_ident(t)}", [], timeout: :infinity)
+    end)
+
+    skipped =
+      Repo.query!("""
+      SELECT count(*) FROM pg_stat_user_tables
+      WHERE schemaname = 'public'
+        AND last_autoanalyze >= now() - ('#{skip_recent} seconds')::interval
+      """).rows
+      |> hd()
+      |> hd()
+
+    %{analyzed: length(tables), skipped: skipped}
+  end
+
+  # Defensive: pg_stat_user_tables returns legitimate identifiers from
+  # the catalog, but we still quote just in case a partition contains a
+  # character the shell would interpret.
+  defp quote_ident(name) when is_binary(name) do
+    escaped = String.replace(name, ~s("), ~s(""))
+    ~s("#{escaped}")
+  end
+
   @doc """
   Derive `surface_refractivity` / `min_refractivity_gradient` /
   `ducting_detected` for every `hrrr_profiles` row where those columns

priv/repo/migrations/20260424204656_tune_autovacuum_for_high_churn.exs

@@ -0,0 +1,66 @@
+defmodule Microwaveprop.Repo.Migrations.TuneAutovacuumForHighChurn do
+  use Ecto.Migration
+
+  @moduledoc """
+  Per-table autovacuum/analyze parameters for the hottest paths.
+
+  Default PG 16 thresholds (`scale_factor 0.2` for vacuum, `0.1` for
+  analyze) are too lazy for a handful of tables we hit frequently:
+
+    * `contacts` — every enrichment tick UPDATEs one or more status
+      columns. 81 k rows × 0.1 analyze scale = 8 k dead tuples before
+      stats refresh; BackfillEnqueueWorker's status-priority ORDER BY
+      then picks a bad plan. Drop to 0.02 / 0.01.
+
+    * `grid_tasks` / `hrrr_fetch_tasks` — small queue tables with a
+      claim/complete cycle that rewrites every row each hour. Defaults
+      mostly keep up, but the claim query's `FOR UPDATE SKIP LOCKED`
+      scan benefits from fresh stats. Set thresholds so vacuum/analyze
+      fire at 50 modified rows instead of ~(default 50) + 20%.
+
+    * `contact_common_volume_radar` — MechanismClassifyWorker UPDATEs
+      bring the analyze stale-tuple counter to ~200 k without ever
+      flipping the default 10 % threshold. Matches the contacts
+      shape at 0.02 / 0.01.
+
+  All other tables keep PG defaults — including `hrrr_profiles`
+  partitions (append-heavy, analyze-fires-naturally on insert) and
+  the Oban-owned tables (`oban_jobs`, `oban_producers`, `oban_peers`),
+  which Oban's own Pruner handles.
+
+  Reversible — `down` resets each back to the cluster-wide default.
+  """
+
+  @high_churn [
+    {"contacts", 0.02, 0.01},
+    {"grid_tasks", 0.05, 0.02},
+    {"hrrr_fetch_tasks", 0.05, 0.02},
+    {"contact_common_volume_radar", 0.02, 0.01}
+  ]
+
+  def up do
+    for {table, vac_scale, analyze_scale} <- @high_churn do
+      execute("""
+      ALTER TABLE #{table} SET (
+        autovacuum_vacuum_scale_factor = #{vac_scale},
+        autovacuum_analyze_scale_factor = #{analyze_scale},
+        autovacuum_vacuum_threshold = 50,
+        autovacuum_analyze_threshold = 50
+      )
+      """)
+    end
+  end
+
+  def down do
+    for {table, _, _} <- @high_churn do
+      execute("""
+      ALTER TABLE #{table} RESET (
+        autovacuum_vacuum_scale_factor,
+        autovacuum_analyze_scale_factor,
+        autovacuum_vacuum_threshold,
+        autovacuum_analyze_threshold
+      )
+      """)
+    end
+  end
+end