Price installed retention and shared source maintenance consistently

[zzylol]

Stacked on #808.

Why

Placement prices could disagree with the installed plan: downstream maintenance needed more panes than priced, and each raw selector inherited the entire workload's sample rate. Distinct logical lookbacks also charged a shared physical Sum producer twice.

What

• Price source retention for downstream maintenance lookbacks.
• Accept optional implementation.source_ingestion_rates evidence by metric, applying fresh rates consistently to retained updates and raw folds. Missing or stale evidence keeps the conservative workload rate.
• Charge generated identical Sum panes within one query once, retaining the longest lookback and every logical read cost. Derived cohorts and provider-measured layouts retain their separate accounting.

Before this PR

A 1m and 10m read sharing one Sum producer charged 68 panes although only 61 were installed. A longer derived consumer could price two source panes when maintenance needed eleven.

After this PR

Those cases price 61 and eleven panes respectively. Explicit source rates replace workload totals for both placement alternatives without assuming label-filter selectivity.

Validation

• Regression failures reproduced before fixes; all 16 lifecycle-placement tests and the focused derived-source regression pass.
• Final-stack validation at Planner a049a3b: formatting, workspace clippy, workspace library tests, all control-plane tests, and all 30 compatibility process tests pass (122.91s). Native candidate retention pricing lives upstream in feat: adopt versioned Planner kernels and typed physical selection #810.
• Independent agent review covered installed-window pricing, freshness, consistent source rates, shared maintenance, and independent mixed-assignment repricing; no blockers remained.

🤖 Generated with Claude Code