docs: fix stacked-levers figure (~3.4d -> ~5.1d) and reconcile r8g to 1.4d

Dawn's Correction 2: stacked levers multiply M, days add — 1.882 x 1.330
x 1.111 = 2.782 -> 5.09d at Td=3.45d (4.93d interaction-aware). The ~3.4d
figure required Td=2.55d, a fast-edge result reported as central. Also
reconciles the kickoff-era r8g '~1.5 days' at line 90 to the fitted 1.4d
used everywhere else. Conclusion unchanged: no infra lever outruns the
curve.

Co-authored-by: npub1qyvc0c5kl4gqv2fd97fsk46tu378sqgy35vc83rvgfwne90sel7s0ed67d <011987e296fd5006292d2f930b574be47c7801048d1983c46c425d3c95f0cffd@buzz.block.builderlab.xyz>
Signed-off-by: npub1qyvc0c5kl4gqv2fd97fsk46tu378sqgy35vc83rvgfwne90sel7s0ed67d <011987e296fd5006292d2f930b574be47c7801048d1983c46c425d3c95f0cffd@buzz.block.builderlab.xyz>

docs/redis-cluster-mode.md

@@ -87,8 +87,8 @@ profile `bb-public-operations-ro`).
   window spanning the onset).
 - Whole-instance CPU ~4%, memory 0.09%, replication lag <1ms. Vertical
   scaling is dead: same-family bigger instances have identical single-core
-  speed; r8g (Graviton4) buys ~25–30% single-thread ≈ 1.5 days at the current
-  doubling rate.
+  speed; r8g (Graviton4) buys ~25–30% single-thread ≈ 1.4 days at the fitted
+  3.45-day doubling (1.33x → Td·log₂(1.33)).
 - **Fit window (measured, P3):** the curve has a sharp onset at 2026-07-21
   ~18:00Z — flat ~0.55% for the preceding five days, then exponential within
   one 15-min bin (subscribed channels 74→1303 in four hours at a pinned
@@ -776,7 +776,7 @@ REDIS_CLUSTER_PRODUCTION_PRIOR_ART.md`, `…_OPEN_SOURCE_PRIOR_ART.md`,
   post-P2** (N=8 post-P2: +4.3d). Any one-time capacity gain of factor M
   buys exactly `Td·log₂(M)` days, so no N reachable this week buys two
   doublings, and every available lever — sharding, r8g (+1.4d), presence
-  reduction, all stacked (~3.4d) — is smaller than the migration's own
+  reduction, all stacked (~5.1d; 4.9d interaction-aware) — is smaller than the migration's own
   duration estimate (14d aggressive / 21d likely / 35d conservative; soft,
   LOC-derived, not a bottom-up plan). **Shard count is therefore not a
   runway purchase, and the honest case for cluster mode is optionality,