docs(plan): strip L3/Vsa10k-BF16 hallucination from Internal/VSA table row

The Internal/VSA dataset row in §17 carried two errors:
1. "L3 cold tier" misread the user's L3-CPU-cache-budget constraint as
   a cognitive-stack layer.
2. "Vsa10k BF16 (20 KB, lossless)" is not a real type — Vsa10k is
   [u64; 157] = ~1.2 KB bit-packed binary (grammar/role_keys.rs:51),
   not a 20 KB bf16 vector. The bf16 variant doesn't exist.

Restore the row to just describe Fingerprint<256> cycle fingerprints
plus NARS truth vectors and braid offsets, which is what the schema
actually carries.

§18 and unified-integration-v1.md's precision note carry the same
cluster of errors and need separate cleanup review.

Author: claude

.claude/plans/callcenter-membrane-v1.md

@@ -819,7 +819,7 @@ git verb.
 | Dataset class | Content | Crossing BBB? | Queryable? |
 |---|---|---|---|
 | **External / scalar** | Arrow RecordBatch rows from `project()` | Yes (IS the BBB output) | Yes — DataFusion, Supabase FDW, n8n subscribe |
-| **Internal / VSA** | `Fingerprint<256>` = `[u64;256]` cycle fingerprints (L4/L5 speed tier, 2 KB/row); NARS truth vectors, braid offsets. L3 cold tier can promote to Vsa10k BF16 (20 KB, lossless) or RaBitQ-quantized Lance columns (zero-copy ANN). | No — never crosses BBB | Yes — DataFusion + VSA UDFs (see § 15) |
+| **Internal / VSA** | `Fingerprint<256>` = `[u64;256]` cycle fingerprints (2 KB/row); NARS truth vectors; braid offsets. | No — never crosses BBB | Yes — DataFusion + VSA UDFs (see § 15) |
 
 ### Parallels with git cold storage