sizecal: ethrex disk cost is ~2x bytesPerSlot (trie + flat rows per slot); docs/CALIBRATION.md missing

Found while reviewing #117 (which fixes the RAM side of #116 but explicitly defers this).

## Problem

`--target-size=350GB` for `--client=ethrex` was heading for **~715 GiB on disk** (extrapolated from the killed run: 249 GiB at 38.1% of Phase 2, plus 60 GiB of Phase-0 spec storage). A local 40 GB fill realized 79 GB — the same ~2×.

Cause: `internal/sizecal/factors.go` documents `bytesPerSlot = 140` as **TRIE-only** (`factors.go:10-14`, "Flat-state rows … are ADDITIONAL and not counted"), but the ethrex writer emits **two rows per slot** — one into `storage_trie_nodes` and one into `storage_flatkeyvalue` (the snap-sync flat-KV layer, both uncompressed) — so ethrex's real per-slot disk cost is ~2× what the planner budgets.

## Constraints

- The constant must stay **identical across clients** for the `cross-client-genesis-root` gate (`factors.go:46-51`): same target → same synthetic entity set → same root. Any fix must keep the *entity counts* identical across clients and only change per-client *disk accounting/expectations* — or raise the global constant for all clients with a coordinated golden-root update.
- The plumbing for a per-client view already exists and is deliberately unused: `SizeApproximator.SlotsForBytes(client, …)` ignores its `client` param (`factors.go:3-8,71,80`).

## Also

`factors.go:13` points at `docs/CALIBRATION.md` for "calibration anchor and per-client flat-state overhead numbers" — **that file does not exist**. Whatever direction this issue lands, CALIBRATION.md should be authored and the per-client multipliers (ethrex ≈ 2×; others per their flat layouts) documented there.

Until fixed: size ethrex runs at ~2× the target on disk (350 GB target ⇒ ~700 GB + Phase-1 sort spill ≈ 0.4× co-resident at peak).

Start with internal/sizecal/factors.go, especially SlotsForBytes and the bytesPerSlot documentation, then inspect the cross-client-genesis-root gate. Add the missing docs/CALIBRATION.md with per-client flat-state overheads and make disk accounting reflect ethrex's two rows per slot without changing entity counts or the shared genesis root.

Written by the indexing model from the issue text.