fix(bgz-tensor): resolve 5 CI-invisible test failures (3 stale tests + 1 slice-bounds bug)

bgz-tensor is excluded from CI (no workflow tests its manifest), so these
reds accumulated unseen on main — surfaced by a local sweep of the
CI-uncovered standalone crates. Diagnosed each as test-stale vs code-bug
before fixing; did not blanket-bump assertions.

Test-stale (code was correct, expected values drifted):
- gamma_calibration: byte_size() is 48 (RoleGamma 36 + CosineGamma 12),
  not 40 — the struct has 8 roles (32B)+phi_scale(4B)=36B; the stale
  '28+12' comment assumed 7 roles.
- hhtl_cache: serialize() writes a 16-byte gamma_meta trailer that the
  test's size formula omitted; added + 16.
- hhtl_d: 0x3C00 is the IEEE-half (F16) bit pattern for 1.0; real BF16
  1.0 is 0x3F80. bf16_to_f32 was correct; the test literal was wrong.

Code-bug (genuine robustness fix, both encode + decode):
- matryoshka encode_row/decode_row panicked when the SVD basis rank is
  lower than the band profile's nominal max (fewer sample rows than
  requested components → a band extends past the available coeffs).
  encode: clamp slice start (was only clamping end); decode: cap the
  per-band component count to the coeff buffer. Symmetric, so the byte
  stream stays in sync; roundtrip + quality assertions pass.

Verified: bgz-tensor 200 passed / 0 failed (was 195/5); fmt clean;
clippy -p bgz-tensor --all-targets -D warnings clean.

Author: claude

crates/bgz-tensor/src/gamma_calibration.rs

@@ -432,7 +432,7 @@ mod tests {
 
     #[test]
     fn calibration_profile_size() {
-        assert_eq!(CalibrationProfile::byte_size(), 40); // 28 + 12
+        assert_eq!(CalibrationProfile::byte_size(), 48); // RoleGamma 36 + CosineGamma 12
     }
 
     #[test]

crates/bgz-tensor/src/hhtl_cache.rs

@@ -630,8 +630,8 @@ mod tests {
         cache.serialize(path).expect("serialize");
 
         let size = std::fs::metadata(path).map(|m| m.len()).unwrap_or(0);
-        // 4 magic + 2 k + 256×34 entries + 256×256×2 distances + 256×256×1 routes + 256×4 radii
-        let expected = 4 + 2 + 256 * 34 + 256 * 256 * 2 + 256 * 256 + 256 * 4;
+        // 4 magic + 2 k + 256×34 entries + 256×256×2 distances + 256×256×1 routes + 256×4 radii + 16 gamma_meta
+        let expected = 4 + 2 + 256 * 34 + 256 * 256 * 2 + 256 * 256 + 256 * 4 + 16;
         assert_eq!(
             size, expected as u64,
             "expected {expected} bytes, got {size}"

crates/bgz-tensor/src/hhtl_d.rs

@@ -553,7 +553,7 @@ mod tests {
 
     #[test]
     fn hhtl_d_entry_roundtrip() {
-        let entry = HhtlDEntry::new(HeelBasin::Gate, 7, 42, true, 0x3C00); // BF16 1.0
+        let entry = HhtlDEntry::new(HeelBasin::Gate, 7, 42, true, 0x3F80); // BF16 1.0 (0x3F80, not the F16 0x3C00)
         let bytes = entry.to_le_bytes();
         let decoded = HhtlDEntry::from_le_bytes(&bytes);
 

crates/bgz-tensor/src/matryoshka.rs

@@ -466,7 +466,7 @@ pub fn encode_row(row: &[f32], basis: &SvdBasis, profile: &BandProfile) -> Matry
 
     for band in &profile.bands {
         let max_val = band.precision.max_val();
-        let band_coeffs = &coeffs[band.start..band.end.min(coeffs.len())];
+        let band_coeffs = &coeffs[band.start.min(coeffs.len())..band.end.min(coeffs.len())];
 
         // Find scale for this band
         let band_max = band_coeffs
@@ -552,7 +552,10 @@ pub fn decode_row(encoded: &MatryoshkaRow, basis: &SvdBasis, profile: &BandProfi
         offset += 2;
         let inv_scale = band_max / max_val as f32;
 
-        let n = band.n_components();
+        // Cap to the coeff buffer: the basis rank can be lower than the
+        // profile's nominal max (e.g. fewer sample rows than requested
+        // components), so a band may extend past the available coeffs.
+        let n = band.n_components().min(coeffs.len().saturating_sub(band.start));
 
         match band.precision {
             BandPrecision::I16 => {