Variable bit allocation per dimension (future, format v1.4)

[ajianaz]

Context

Quantization Beyond Uniform Bit Allocation: allocate more bits to high-variance dims, fewer to low-variance ones, at the same total budget. For rotated Gaussian-ish dims (post-RHDH) the variance is uniform by construction — so this likely matters only in combination with #24 (Matryoshka truncation leaves non-uniform effective variance) or un-rotated variants.

Why parked

• Post-RHDH uniformity may nullify the gain — needs mmap view: serve an index from disk without loading into RAM #25's benchmark baseline first.
• Format change (per-dim bit widths) — do at most ONE more format migration after v1.2 (Residual quantization (TurboQuant-style second pass) — recall at same bit width #23).

Acceptance (when picked up)

• Simulation showing expected recall gain on our data before any format work
• Format v1.3 design RFC

Dependencies

Evidence-gated on #25. Future.

[ajianaz]

Closing: premise is largely invalidated by the RHDH design

Per-dimension bit allocation only pays when per-dimension reconstruction error dominates. But vecq's pipeline deliberately equalizes per-dimension variance: the seeded RHDH rotation + FWHT spreads every vector's energy so rotated coordinates are approximately N(0,1) — that is the flatness property the whole quantizer relies on (README "How it works" step 1). Allocating more bits to "high-variance" rotated dimensions would chase noise the rotation just manufactured.

Where a coarse version of this idea does make sense, it already shipped:

• Pre-rotation bit allocation = choosing which source dims matter → that's Matryoshka working_dim (v1.2), which truncates low-value dims entirely (a stronger allocation than bits).
• More precision where needed → configurable uniform width 4/5/6-bit (feat(core): parameterize quantizer bit-width (5/6-bit) — dominates residual mode on size AND speed #39) and the residual second pass (Residual quantization (TurboQuant-style second pass) — recall at same bit width #23) cover the precision/recall trade-space without breaking the training-free pitch.

Adding per-dim allocation today would cost: per-dim variance estimation (a training pass — against the project thesis), a format change (v1.6 layout + migration), and non-uniform packing that breaks the fixed-position view-friendly layout (#25) — for expected gains near zero on rotated dims.

Reopen with data if a workload shows rotated-dim variance skew ≥ ~2x after RHDH — that would falsify this analysis.