Problem

FixedToLogQuantized clamps its input to [-32768, 32767] (i.e. |value| < 0.5 in Q16.16) before a table lookup:

IF fp < -32768 THEN fp = -32768
ELSEIF fp > 32767 THEN fp = 32767
END IF
lq.packed_value = QuantFixedLookup(fp)

Two consequences, both bad for real weights:

1. Every weight with |value| >= 0.5 is forced to 0.5. Round-tripping 1.0, 2.0, 5.0 all yield 0.5 — even though the log format (exponent in [0,15], mantissa in [0,15]) already represents magnitudes up to ~120.0. The clamp throws away almost the entire representable range.
2. Many small weights collapse to exactly 0. On a realistic GPT-2 weight distribution (~N(0, 0.05)), ~3.7% of weights quantize to 0.

Fix

Encode directly with a closed form — no clamp, no lookup table — using the same packing ((exponent << EXPONENT_SHIFT) | mantissa) that InitQuantizationTables and LogQuantizedToFixed already dequantize. It's O(1): a bounded MSB scan (<= ~20 iterations), not the 511-entry nearest-code search one might otherwise reach for.

DIM a AS LONG = ABS(CLNG(fp))
DIM t AS LONG = a, p AS INTEGER = 0
WHILE t > 1 : t = t SHR 1 : p += 1 : WEND
DIM e AS INTEGER = p - 7 : IF e < 0 THEN e = 0
IF e > 15 THEN e = 15
DIM m AS INTEGER = (a + (CLNG(1) SHL (e + 3))) SHR (e + 4)  ' round-to-nearest
IF m > 15 THEN m = 15
DIM code AS INTEGER = (e SHL EXPONENT_SHIFT) OR m
IF fp < 0 THEN code = -code

Evidence (measured in the real DOS runtime)

Cross-compiled with DJGPP FreeBASIC and run under DPMI (dosbox-x); numbers agree bit-for-bit with a 32-bit Linux fbc reference.

Accuracy — RMS reconstruction error:

→ 19.6x lower error on realistic weights, within 6% of the exhaustive optimum. Original collapses 1867/50000 weights to 0; this fix collapses none.

Cost — per call, cycles=max:

→ constant-time; ~36x faster than a nearest-code search at matching accuracy, a small constant over the (incorrect) original.

Checks

• Full src/main_prod.bas cross-compiles to DOS clean with this change.
• python3 -m unittest discover tests → 108 passed; git diff --check clean; verify_workspace_tracking.py passes.
• Diff is one function in src/quantization.bas (+31 / -11).

Note

QuantFixedLookup (and its fill loop in InitQuantizationTables) is now unused by the encode path. I left it in place to keep this diff to one function — happy to remove it as a follow-up if you'd prefer.