docs(research): GPTQ max-exactness fix -- partial, closes ~35% of the RTN gap

[gHashTag]

Runs the fix proposed but not executed in THEOREM_2026-08-09.md: protect each group's per-row maximal element from GPTQ's sequential compensation (equivalent to processing it first, since an exactly-quantised column propagates zero error).

• max-exactness restored: 95.3% -> 0.4% blocks with an inexact row-max
• perplexity: GPTQ-original 17.7047 -> GPTQ-fixed 17.5860 (-0.1186, same run)
• still worse than RTN 4-bit (17.3667) by +0.2194 -- gptq_gate.py's pass condition remains unmet
• reproduction note: this environment's GPTQ-original (17.7047) differs from THEOREM's (17.7846) by 0.08; all comparisons here are within-run to control for that

Conclusion: max-exactness destruction explains part (~35%) of GPTQ's gap to RTN under block-max scaling, not all of it. Residual mechanism unidentified -- open question, not run further (5-bit / promote-only / cross-model out of scope for this PR).

Adds research/block/gptq_maxexact_fix.py, research/block/GPTQ_MAXEXACT_FIX_2026-08-11.md.

[gHashTag]

Reviewer bee Z: sound, merging.

Evidence (head c072c38):

• The PR is additive: 2 new files under research/block/, and the merge into main is clean. Nothing on main covers it: the only max-exactness text is the proposal in THEOREM_2026-08-09.md.
• gptq_maxexact_fix.py compiles. Its imports are numpy, torch, transformers and pyarrow only, with no network, subprocess or eval calls. The hardcoded weights path follows the same convention as its siblings (e.g. block_tnf.py).
• The arithmetic in the doc re-checks:
  • 3,162,231/3,317,760 = 95.3%
  • 12,211/3,317,760 = 0.37%
  • gap 17.7047−17.3667 = 0.338 → 0.2193, i.e. −35.1%
• The conclusion is stated honestly: the fix is partial, the gate is not cleared, and the residual is unexplained.

Nit, not blocking: the script's own repro_ok check (|GPTQ-orig − 17.7846| < 0.02) is false in the environment that produced these numbers (Δ 0.080). The doc discloses this and keeps every comparison within one run, which is the right control.